Split (1)

train · 2.34k rows

title stringlengths 28 135	abstract stringlengths 0 12k	introduction stringlengths 0 12k
Zhang_TokenHPE_Learning_Orientation_Tokens_for_Efficient_Head_Pose_Estimation_via_CVPR_2023	Abstract Head pose estimation (HPE) has been widely used in the fields of human machine interaction, self-driving, and attention estimation. However, existing methods cannot deal with extreme head pose randomness and serious oc- clusions. To address these challenges, we identify three cues from head images, namely, neighborhood similari- ties, significant facial changes, and critical minority rela- tionships. To leverage the observed findings, we propose a novel critical minority relationship-aware method based on the Transformer architecture in which the facial part relationships can be learned. Specifically, we design sev- eral orientation tokens to explicitly encode the basic ori- entation regions. Meanwhile, a novel token guide multi- loss function is designed to guide the orientation tokens as they learn the desired regional similarities and relation- ships. We evaluate the proposed method on three chal- lenging benchmark HPE datasets. Experiments show that our method achieves better performance compared with state-of-the-art methods. Our code is publicly available at https://github.com/zc2023/TokenHPE .	1. Introduction Head pose estimation (HPE) is a popular research area in computer vision and has been widely applied to driver assis- tance [29], human–computer interaction [36], virtual reality [22], and attention detection [5]. In recent years, HPE has been actively studied and the accuracy has been consider- ably improved in terms of utilizing extra facial landmark in- formation [2,20], extra RGB-depth information [13,26–28], extra temporal information [16], stage-wise regression strat- egy [42], multi-task learning [1, 38], and alternative param- eterization of orientation [3,15,18,19,24]. Currently, many methods focus on the representation of the head pose ori- *: Corresponding author (Hai Liu, Youfu Li) (i) (ii) (iii)Neighbored orientations similarity and significant facial part changesChallenge of facial part missing vs. Finding of critical minority relationships Missing parts Critical minority partsFigure 1. Left part: Missing facial parts and our finding on critical minority relationships. Although some of the facial parts are miss- ing or occluded (marked with a red rectangle), the pose orientation still can be inferred from the existing critical minority facial parts (marked with a green circle). Right part: different head orientation images in a panoramic overview. The rectangular boxes highlight several significant facial changes, such as i) appearance of the eye on one side, ii) appearance of the nostril, and iii) overlapping of the nose and mouth. The circled areas show some regions in which the facial part features are similar. entation and have achieved impressive performance, but the intrinsic facial part relationships are usually neglected. A possible reason is that these relationships are difficult to learn by existing CNN architectures. However, in some challenging scenarios, as shown in the left part of Fig. 1, many remarkable facial parts are missing. Consequently, the remaining facial parts and their geometric relationships must be leveraged to achieve robust and high-accuracy pre- diction. Therefore, how to leverage the facial part relation- ships for high-accuracy HPE is an attractive research topic. To begin with, we firstly identify three implicit facial part relationships in head poses. First, a local similarity in spe- cific spatial orientation exists. Inside the circled region in Fig. 1, the facial appearances are similar. Second, sev- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 8897 eral significant facial part changes are observed in specific orientations. For example, in Fig. 1, the two circled fa- cial regions can be distinguished by a significant facial part change, which is the appearance of the right eye. Third, critical minority relationships of facial parts exist, and they can determine the orientation of a head pose despite pos- sible occlusions. As Fig. 1 shows, if a person’s mouth is occluded, the head pose can be determined by the geomet- ric spatial relationships of the eyes, nose, and the outline of the face. In these scenarios, the remaining minor facial parts and their relationships are crucial for high-accuracy HPE. Given the aforementioned facial part relationships, the question is how to design a model that can utilize this heuristic knowledge. The traditional CNN architecture can- not easily learn these relationships. In contrast, the Trans- former architecture can effectively address this drawback of CNN. Recently, Vision Transformer (ViT) [11] emerged as a new choice for various computer vision tasks. The Trans- former architecture is known for its extraordinary ability to learn long-distance, high-level relationships between image patches. Therefore, using Transformer to learn the rela- tionships among critical minority facial parts is reasonable. Moreover, the basic orientation regions can be well repre- sented by learnable tokens in Transformer. Inspired by the three findings and Transformer’s prop- erties, we propose TokenHPE, a method that can discover and leverage facial part relationships and regional similari- ties via the Transformer architecture. The proposed method can discover facial part geometric relationships via self- attention among visual tokens, and the orientation tokens can encode the characteristics of the basic orientation re- gions. The latent relationships between visual and orienta- tion tokens can be learned from large HPE datasets. Then, the learned information is encoded into the orientation to- kens, which can be visualized by vector similarities. In addition, a special token guide multi-loss function is con- structed to help the orientation token learn the general in- formation. Our main contributions can be summarized as follows: (1) Three findings are derived on head images, including facial part relationships and neighborhood orientation simi- larities. Furthermore, to leverage our findings and cope with challenging scenarios, a novel token-learning model based on Transformer for HPE is presented for the first time. (2) We find that the head pose panoramic overview can be partitioned into several basic regions according to the orientation characteristics. The same number of learnable orientation tokens are utilized to encode this general infor- mation. Moreover, a novel token guide multi-loss function is designed to train the model. (3) We conduct experiments on three widely used HPE datasets. TokenHPE achieves state-of-the-art performance with a novel token-learning concept compared with its ex-isting CNN-based counterparts. Abundant visualizations are also provided to illustrate the effectiveness of the pro- posed orientation tokens.
Yao_Hi-LASSIE_High-Fidelity_Articulated_Shape_and_Skeleton_Discovery_From_Sparse_Image_CVPR_2023	Abstract Automatically estimating 3D skeleton, shape, camera viewpoints, and part articulation from sparse in-the-wild image ensembles is a severely under-constrained and chal- lenging problem. Most prior methods rely on large-scale image datasets, dense temporal correspondence, or human annotations like camera pose, 2D keypoints, and shape tem- plates. We propose Hi-LASSIE, which performs 3D articu- lated reconstruction from only 20-30 online images in the wild without any user-defined shape or skeleton templates. We follow the recent work of LASSIE that tackles a similar problem setting and make two significant advances. First, instead of relying on a manually annotated 3D skeleton, we automatically estimate a class-specific skeleton from the selected reference image. Second, we improve the shape reconstructions with novel instance-specific optimization strategies that allow reconstructions to faithful fit on each instance while preserving the class-specific priors learned *Work done as a student researcher at Google.across all images. Experiments on in-the-wild image en- sembles show that Hi-LASSIE obtains higher fidelity state- of-the-art 3D reconstructions despite requiring minimum user input. Project page: chhankyao.github.io/ hi-lassie/	1. Introduction 3D assets of articulated animals enable numerous appli- cations in games, movies, AR/VR, etc. However, building high-fidelity 3D models of articulated shapes like animal bodies is labor intensive either via manual creation or 3D scanning. Recent advances in deep learning and 3D rep- resentations have significantly improved the quality of 3D reconstruction from images. Much of the success depends on the availability of either rich 3D annotations or multi- view captures, both of which are not always available in a real-world scenario. A more practical and scalable alter- native is automatic 3D reconstruction from online images as it is straightforward to obtain image ensembles of any animal category ( e.g., image search results). In this work, This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 4853 Reference image …3D skeleton3D shape3D shape with surface mapsLASSIEHi-LASSIE (ours)SMALR, 3DSafariCSM, A-CSM Figure 2. User inputs across different techniques for articulated animal reconstruction. Contrary to prior methods that leverage detailed 3D shapes or skeletons, Hi-LASSIE only requires the user to select a reference image where most animal body parts are visible. we tackle a practical problem setting introduced in a recent work, LASSIE [38], where the aim is to automatically esti- mate articulated 3D shapes with only a few (20-30) in-the- wild images of an animal species, without any image-level 2D or 3D annotations. This problem is highly under-constrained and challeng- ing due to a multitude of variations within image ensembles. In-the-wild images usually have diverse backgrounds, light- ing, and camera viewpoints. Moreover, different animal in- stances can have distinct 2D appearances due to pose ar- ticulations, shape variations, and surface texture variations (skin colors, patterns, and lighting). As shown in Fig. 2, early approaches in this space try to simplify the problem with some user-defined 3D templates, hurting the general- ization of those techniques to classes where such templates are not always readily available. In addition, most of these methods (except LASSIE [38]) assume either large-scale training images of each animal species [16, 17, 43] or per- image 2D keypoint annotations [44], which also limits their scalability. In this work, we propose a more practical technique that does not require any 3D shape or skeleton templates. In- stead, as illustrated in Fig. 2, the user simply has to se- lect a reference image in the ensemble where all the ani- mal parts are visible. We achieve this by providing two key technical advances over LASSIE: 1) 3D skeleton discovery and 2) instance-specific optimization strategies. Our over- all framework, named Hi-LASSIE, can produce Higher- fidelity articulated shapes than LASSIE [38] while requir- ing minimal human input. Our key insight is to exploit the 2D part-level correspondences for 3D skeleton discov- ery. Recent works [1, 32] observe that the deep features ex- tracted from a self-supervised vision transformer (ViT) [11] like DINO-ViT [7] can provide good co-part segmentation across images. We further exploit such features to rea- son about part visibility and their symmetry. At a high level, we first obtain a 2D skeleton using the animal silhou- ette and part clusters [1] obtained from DINO features [7]. We then uplift this 2D skeleton into 3D using symmet- ric part information that is present in the deep DINO fea- tures. Fig. 1 shows the skeleton for zebra images discov-ered by Hi-LASSIE. Similar to LASSIE [38], we leverage 3D part priors (learned from and shared across instances) and the discovered 3D skeleton to regularize the articu- lated shape learning. Furthermore, we design three novel modules to increase the quality of output shapes: 1) High- resolution optimization by zooming in on individual parts, 2) Surface feature MLPs to densely supervise the neural part surface learning, and 3) Frequency-based decomposi- tion of part surfaces for shared and instance-specific com- ponents. Note that Hi-LASSIE can generalize to diverse animal species easily as it does not require any image anno- tations or category-specific templates. We conduct extensive experiments on the Pascal-Part [8] and LASSIE [38] image ensembles, which contain in-the- wild images of various animal species like horse, elephant, and penguin. Compared with LASSIE and other baselines, we achieve higher reconstruction accuracy in terms of key- point transfer, part transfer, and 2D IOU metrics. Quali- tatively, Hi-LASSIE reconstructions show considerable im- provement on 3D geometric and texture details as well as faithfulness to input images. Finally, we demonstrate sev- eral applications like animation and motion re-targeting en- abled by our 3D part representation. Fig. 1 (right) shows some Hi-LASSIE 3D reconstructions for different animal species. The main contributions of this work are: • To our best knowledge, Hi-LASSIE is the first approach to discover 3D skeletons of articulated animal bodies from in-the-wild image ensembles without using any image-level annotations. We show that the discovered 3D skeleton can faithfully fit all instances in the same class and effectively regularize the 3D shape optimization. • Hi-LASSIE includes several novel optimization strategies that makes the output shapes richer in 3D details and more faithful to each image instance than prior methods. • Extensive results on multiple animal classes and datasets demonstrate the state-of-the-art performance of Hi-LASSIE while requiring less user inputs than prior works.
Zhang_Decoupling_MaxLogit_for_Out-of-Distribution_Detection_CVPR_2023	Abstract In machine learning, it is often observed that standard training outputs anomalously high confidence for both in- distribution (ID) and out-of-distribution (OOD) data. Thus, the ability to detect OOD samples is critical to the model de- ployment. An essential step for OOD detection is post-hoc scoring. MaxLogit is one of the simplest scoring functions which uses the maximum logits as OOD score. To provide a new viewpoint to study the logit-based scoring function, we reformulate the logit into cosine similarity and logit norm and propose to use MaxCosine and MaxNorm. We empir- ically find that MaxCosine is a core factor in the effective- ness of MaxLogit. And the performance of MaxLogit is en- cumbered by MaxNorm. To tackle the problem, we propose the Decoupling MaxLogit (DML) for flexibility to balance MaxCosine and MaxNorm. To further embody the core of our method, we extend DML to DML+ based on the new insights that fewer hard samples and compact feature space are the key components to make logit-based methods effec- tive. We demonstrate the effectiveness of our logit-based OOD detection methods on CIFAR-10, CIFAR-100 and Im- ageNet and establish state-of-the-art performance.	1. Introduction In real-world applications, the closed-world assumption does not always hold where all the classes in the test phase would be available in the training phase. Out-of-distribution (OOD) detection [11] is a natural and challenging setting, and there is an open space containing outliers not belong- ing to any training classes. When the model is deployed in practice, OOD data often come from the open world [17]. Thus, it is crucial for a trustworthy model to not only pro- duce accurate predictions on in-distribution (ID) data, but also distinguish the OOD data and reject them. However, the machine-learning model easily produces over-confident *Equal contribution, co-first author; also with Nat. Key Lab of MSIIPT. Correspondence to Xiang Xiang ([email protected]) 50 60 70 80 90 100 CIFAR-100 Mean AUROC818487909396ImageNet Mean AUROCGradNormEnergy MSPMaxLogit ViMODINDMLDML+(a) AUROC of different methods Textures SVHN LSUN-C LSUN-R iSUN Places365020406080100FPR95 scoreOurs SOTA w/o training SOTA w/ training (b) FPR95 of SOTA methods Figure 1. AUROC and FPR95 (in percentage) of previous OOD detection methods on ImageNet and CIFAR-100. (a) shows the AUROC (higher is better) of our methods (orange pentagons) and other methods (blue rectangles). (b) shows the FPR95 (lower is better) of our methods and SOTA methods w/ (LogitNorm [37]) and w/o (ViM [36]) training on CIFAR-100. wrong predictions on OOD data [25]. For instance, a model may wrongly detect the zebra as a horse with high confi- dence when the zebra is not in the training set. A key of the OOD detection algorithm is a scoring func- tion that maps the input to the OOD score, indicating to what extent the sample is an OOD sample. Various scor- ing functions have been proposed to seek the properties that better distinguish OOD samples. The OOD score is calculated mainly from the output of the model, includ- ing features [20, 30,36], logits [10, 11,23]. For example, MSP [11] uses the maximum Softmax probabilities, and MaxLogit [10] uses the maximum logits as the OOD score. MaxLogit and MSP are two simplest scoring functions that do not require extra computational costs. In contrast, other methods require extra storage [30], or extra compu- tational cost [14]. However, the logit-based methods MSP and MaxLogit are not state-of-the-art (SOTA). Intuitively, the simple logit-based method could achieve comparable performance as other complex scoring methods, because logit contains high-level semantic information. We hypoth- esize some underlying reasons limit the performance. To revitalize the simple logit-based method, we start the work by analyzing the reasons which cause the performance gap between MSP and MaxLogit. The gap may be due to This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 3388 the softmax operation normalizing out much feature norm information of the logits. To delve into the effect of feature norm, we divide the logit into two parts: (1) the cosine sim- ilarity between the features and classifier; (2) the feature norm. We discard the classifier weight norm because the norm is identical after the model coverage [27]. We use the top value of the two as the OOD score, named MaxCosine and MaxNorm. Therefore, MaxLogit is a coupled form. We find that MaxCosine outperforms MaxLogit with the same model and MaxNorm performs much worse than MaxLogit. Thus, MaxLogit (1) is encumbered by MaxNorm, (2) suppresses the effectiveness of MaxCo- sine, and (3) restricts the flexibility to balance MaxCosine and MaxNorm. The three problems are the bottleneck of MaxLogit. To tackle the problem, we propose Decoupling MaxLogit (DML) for flexibility to balance MaxCosine and MaxNorm. DML decouples the MaxCosine from the equal coefficient with MaxNorm by replacing it with a constant. The decoupling method solves the second and third problems but still leaves the first problem unsolved. Al- though MaxNorm helps DML to outperform MaxCosine, the improvement is marginal due to the low performance of MaxNorm. Therefore, we study the role of model train- ing and show that a simple modification to standard train- ing could significantly boost MaxNorm and MaxCosine for OOD detection. Specifically, a feature space with fewer hard samples benefits MaxCosine and a compact feature space benefits MaxNorm. Also, the normalized feature and classifier are the key to the success of the logit-based meth- ods. These findings are not discussed in prior works. We extend DML to DML+ based on the above new insights to further boost the DML performance as shown in Fig. 1. We summarize our contributions as follows. • To overcome the limitations of MaxLogit, we propose a post-hoc scoring method DML, which decouples MaxLogit for flexibility to balance MaxCosine and MaxNorm. DML outperforms MaxLogit and achieves comparable performance with SOTA methods. • We offer new insights into the key components to make MaxCosine, MaxNorm and DML effective, including replacing the standard linear classifier with a cosine classifier and different training losses. The findings are supported by empirical results and theoretical analysis. We also prove that the findings could greatly boost the performance of existing OOD scoring methods. • Based on the insights, we extend DML to DML+ which changes the standard training. Significant im- provements on CIFAR and ImageNet have shown its effectiveness.
Yang_BiCro_Noisy_Correspondence_Rectification_for_Multi-Modality_Data_via_Bi-Directional_Cross-Modal_CVPR_2023	Abstract As one of the most fundamental techniques in multi- modal learning, cross-modal matching aims to project various sensory modalities into a shared feature space. To achieve this, massive and correctly aligned data pairs are required for model training. However, unlike unimodal datasets, multimodal datasets are extremely harder to collect and annotate precisely. As an alternative, the co-occurred data pairs (e.g., image-text pairs) collected from the Internet have been widely exploited in the area. Unfortunately, the cheaply collected dataset unavoidably contains many mismatched data pairs, which have been proven to be harmful to the model’s performance. To address this, we propose a general framework called BiCro (Bidirectional Cross-modal similarity consistency), which can be easily integrated into existing cross-modal matching models and improve their robustness against noisy data. Specifically, BiCro aims to estimate soft labels for noisy data pairs to reflect their true correspondence degree. The basic idea of BiCro is motivated by that – taking image-text matching as an example – similar images should have similar textual descriptions and vice versa. Then the consistency of these two similarities can be recast as the estimated soft labels to train the matching model. The ex- periments on three popular cross-modal matching datasets demonstrate that our method significantly improves the noise-robustness of various matching models, and surpass the state-of-the-art by a clear margin.The code is available athttps://github.com/xu5zhao/BiCro .	1. Introduction As a key step towards general intelligence, multimodal learning aims to endow an agent with the ability to extract and understand information from various sensory modali- ties. One of the most fundamental techniques in multimodal *Corresponding Author([email protected]). Soccer player is beaten to the ball by soccer player.I1T1 I2Look at him, like it’s no work at all. He is doing some works.T2˜I˜Tclean pairsnoisy pairsFigure 1. Illustration of the Bidirectional Cross-modal similarity consistency (BiCro) . In an image-text shared feature space, as- sume (I1, T1)and(I2, T2)are two clean positive data pairs. The image ˜Iis very close to the image I1, but their corresponding texts ˜TandT1are far away from each other. Similarly, the text ˜Tis very close to the text T2while their corresponding images ˜IandI2are irrelevant. Therefore the data pair ( ˜I,˜T) is possibly noisy (mis- matched). learning is cross-modal matching. Cross-modal matching tries to project different modalities into a shared feature space, and it has been successfully applied to many areas, e.g., image/video captioning [1, 21, 23], cross-modal hash- ing [14, 18], and visual question answering [15, 58]. Like many other tasks in deep learning, cross-modal matching also requires massive and high-quality labeled training data. But even worse, the multimodal datasets ( e.g., in image-text matching) are significantly harder to be man- ually annotated than those unimodal datasets ( e.g., in image classification). This is because the textual description of an image is very subjective, while the categorical label of an image is easier to be determined. Alternatively, many works collect the co-occurred image-text pairs from the Internet to train the cross-modal matching model, e.g., CLIP [32] was trained over 400 million image-text pairs collected from the Internet. Furthermore, one of the largest public datasets in cross-modal matching, Conceptual Caption [36], was also This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 19883 harvested automatically from the Internet. Such cheap data collection methods inevitably introduce noise into the col- lected data, e.g., the Conceptual Caption dataset [36] con- tains 3%∼20% mismatched image-text pairs. The noisy data pairs undoubtedly hurt the generalization of the cross- modal matching models due to the memorization effect of deep neural networks [55]. Different from the noisy labels in classification, which refer to categorical annotation errors, the noisy labels in cross-modal matching refer to alignment errors in paired data. Therefore most existing noise-robust learning meth- ods developed for classification cannot be directly applied to the cross-modal matching problem. Generally, all the data pairs in a noisily-collected cross-modal dataset can be categorized into the following three types based on their noise level: well-matched ,weakly-matched , and mis- matched . The well-matched data pairs can be treated as clean examples, and the weakly-matched andmismatched data pairs are noisy examples since they are all labeled as positive ( y= 1) in the dataset. We illustrate some noisy data pairs, including both weakly-matched andmismatched pairs in Fig. 2. The key challenge in noise-robust cross- modal matching is how to estimate accurate soft correspon- dence labels for those noisy data pairs. The estimated soft label is expected to be able to depict the true correspon- dence degree between different modalities in a data pair. In this work, we propose a general framework for soft correspondence label estimation given only noisily- collected data pairs. Our framework is based on the Bidirec- tional Cross-modal similarity consistency (BiCro) , which inherently exists in paired data. An intuitive explanation of the BiCro is – taking image-text matching task as an example – similar images should have similar textual de- scriptions and vice versa. In other words, if two images are highly similar but their corresponding texts are irrele- vant, there possibly exists image-text mismatch issue (see Fig. 1). Specifically, we first select a set of clean positive data pairs as anchor points out of the given noisy data. This step is achieved by modeling the per-sample loss distribu- tion of the dataset using a Beta-Mixture-Model (BMM) and selecting those samples with a high probability of being clean. Then, with the help of the selected anchor points, we compute the Bidirectional Cross-modal similarity con- sistency (BiCro) scores for every noisy data pair as their soft correspondence labels. Finally, the estimated soft cor- respondence labels are recast as soft margins to train the cross-modal matching model in a co-teaching manner [11] to avoid the self-sample-selection error accumulation prob- lem. Before delving into details, we summarize our main con- tributions below: • This paper tackles a widely-exist but rarely-explored problem in cross-modal matching, i.e.,the noisy cor-respondence issue. We identify and explore the key challenge of noise-robust cross-modal learning: how to rectify the binary noisy correspondence labels into more accurate soft correspondence labels for noisy data pairs. • We propose a general framework called Bidirectional Cross-modal similarity consistency (BiCro) for soft correspondence label estimation given only noisy data pairs. The BiCro is motivated by a rational assumption that – taking the image-text matching as an example –similar images should have similar textual descrip- tions and vice versa. The BiCro can be easily adapted to any cross-modal matching models to improve their noise robustness. • To compute the BiCro score for noisy data pairs, we propose to first select some clean positive data pairs asanchor points out of the noisy dataset by mod- eling the per-sample loss distribution of the dataset through a Beta-Mixture-Model (BMM) . Then a set of examples with high clean possibilities can be se- lected. The BMM has a better modeling ability on the skewed clean data loss distribution compared to the Gaussian-Mixture-Model (GMM) used in the previous method [12]. • Extensive experiments on three cross-modal match- ing datasets, including both synthetic and real-world noisy labels, demonstrate that our method surpasses the state-of-the-art by a large margin. The visualiza- tion results also indicate that our estimated soft corre- spondence labels are highly-aligned with the correla- tion degree between different modalities in data pairs.
Zhang_Efficient_RGB-T_Tracking_via_Cross-Modality_Distillation_CVPR_2023	Abstract Most current RGB-T trackers adopt a two-stream struc- ture to extract unimodal RGB and thermal features and complex fusion strategies to achieve multi-modal featurefusion, which require a huge number of parameters, thus hindering their real-life applications. On the other hand,a compact RGB-T tracker may be computationally efﬁ-cient but encounter non-negligible performance degrada-tion, due to the weakening of feature representation abil-ity. To remedy this situation, a cross-modality distillationframework is presented to bridge the performance gap be-tween a compact tracker and a powerful tracker . Speciﬁ-cally, a speciﬁc-common feature distillation module is pro-posed to transform the modality-common information aswell as the modality-speciﬁc information from a deeper two-stream network to a shallower single-stream network. Inaddition, a multi-path selection distillation module is pro-posed to instruct a simple fusion module to learn more ac-curate multi-modal information from a well-designed fusion mechanism by using multiple paths. We validate the effec- tiveness of our method with extensive experiments on threeRGB-T benchmarks, which achieves state-of-the-art perfor- mance but consumes much less computational resources.	1. Introduction RGB-T tracking is the task of estimating the state of an arbitrary target in each frame of an RGB-T video se- quence [ 35]. Due to the affordability of thermal infrared (TIR) sensors, RGB-T tracking draws more and more re-search interest. As shown in Fig. 1(a), most existing RGB-T tracking models ﬁrst adopt a two-stream structure to extract multi-level unimodal RGB and TIR features, respectively, andthen employ elaborate-designed multi-modal feature fusion modules to exploit complementary information within the *Corresponding author.Complex Fusion ConV ConV ConV ConV ConV ConVComplex FusionComplex FusionTarget Estimation Simple Fusion ConV ConV ConV ConV ConV ConVSimple FusionSimple FusionTarget EstimationShared Shared Target EstimationTeacher: Two-stream Feature Extractor & Complex Fusion Student: Single-stream Feature Extractor & Simple FusionCross-Modality Distillation(a) Two-stream Structure (b) Single-stream Structure (c) Our proposed methodConV ConVConv block in RGB feature extractor Conv block in TIR feature extractor ConV Conv block in single-stream feature extractor Fusion Fusion Module RGB data flow TIR data flow TIR & RGB data flow Figure 1. Architectures of different RGB-T tracking models. (a) Two-stream structure. (b) Single-stream structure. (c) Our pro-posed method. multi-modal data. Finally, they deduce the target state, of- ten represented by a bounding box, from the fused features.Although great progress has been made, these powerfulRGB-T tracking models usually require high computationalcosts and large model sizes to handle the information of twomodalities in the stages of unimodal feature extraction andmulti-modal feature fusion. There are two straightforward solutions to tackle the complexity and efﬁciency issues. One is to adopt a single- stream feature extractor with fewer convolutional layers, and the other is to employ simpler multi-modal feature fu-sion modules, as shown in Fig. 1(b). Although such com- pact models can reduce computational complexity, they in-evitably bring non-negligible performance degradation due to the weakening of unimodal feature representation ability and multi-modal complementary information exploration ability. For instance, a powerful RGB-T tracker [ 35] with a two-stream structure and complicated multi-modal feature fusion modules suffers from severe performance degrada- tion after the above model simpliﬁcation operations (84.4%precision rate vs78.1% precision rate on RGBT234 dataset [11]), as shown in Fig. 2. Now, the research question becomes: can we shrink the RGB-T tracker without sacriﬁcing performance? This paper answers this question using knowledge distillation,which allows a compact model to obtain a similar abil-ity of a complex model at little cost. We call this com- plex but powerful model the teacher model, and call this This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 5404 77.578.579.580.581.582.583.584.585.5 15 17 19 21 23 25 27 29 31 33Student1: Student + SCFD Student2: Student + SCFD + MPSD Student3: Student + SCFD + MPSD+HFRDStudent1Student2Student3 StudentTeacher 10M 20M40M Model Speed (FPS)Precision Rate (%) Figure 2. Experimental results of different RGB-T tracking struc- tures on RGBT234 dataset [ 11]. Teacher denotes the two-stream structure with complex fusion modules. Student denotes a single- stream structure with simple fusion operations. The teacher modelemploys ResNet50 [ 7] for feature extraction and fusion modules in [35] for multi-modal feature fusion, respectively. The student model employs ResNet18 [ 7] for feature extraction and concate- nation for multi-modal feature fusion, respectively. compact model the student model. Although some works [15,19,20,29] have made considerable progress on knowl- edge distillation in multi-modal tasks, they fail to conduct adeep investigation on the huge feature differences betweenteacher and student in the unimodal feature extraction stageas well as in the multi-modal feature fusion stage, therebyresulting in suboptimal efﬁciency of the knowledge trans- formation. For that, a novel teacher-student knowledge dis-tillation training framework, named Cross-Modality Dis- tillation (CMD), is proposed to elaborately guide efﬁcientimitation from three stages: unimodal feature extraction,multi-modal feature fusion and target estimate estimation, as shown in Fig. 1(c). Speciﬁcally, in the stage of unimodal feature extraction, as pointed out by many previous works [ 9,17], the shal- lower layers of unimodal features usually contain abun- dant low-level spatial details, which are usually modality-dependent. Differently, the deeper layers of unimodal fea- tures often contain many high-level semantic cues, which tend to be strongly modality-consistent. The student modeluses a compact single-stream network to extract both RGBfeatures and TIR features, which not only lacks the ability toextract modality-speciﬁc information in the shallower lay-ers, but also insufﬁciently explores the modality-commoninformation in the deeper layers. These interesting observa-tions inspire us to design a Speciﬁc-common Feature Dis-tillation (SCFD) module, which transforms the modality-speciﬁc information as well as the modality-common in-formation from a two-stream deeper network to a single- stream shallower network. Second, in the stage of multi-modal feature fusion, the complex multi-modal feature fusion modules in the teacher model show great advantages in various scenarios, whilethe simple fusion strategies in the student model are usu- ally effective in some speciﬁc scenarios. It is difﬁcult for astudent model with a single simple fusion strategy to learn more effective complementary information mining capabil- ities from a complex teacher model due to the large fea- ture differences. Therefore, we design a fusion module withmultiple simple fusion strategies in the student model, de- noted as Multi-path Selection Distillation (MPSD) module.In the process of learning from the teacher model, the stu-dent model can adaptively combine different types of fusion features to make up for the lack of complementary informa- tion mining capabilities of a single simple fusion strategy. Finally, in the stage of target state estimation, with the weakening of the feature representation ability of the stu- dent model, the discriminative ability of the tracker fordistractors is also reduced. For that, we further present a Hard-focused Response Distillation (HFRD) module to im- prove the student model’s discriminative ability by alleviat-ing the problem of data imbalance between the targets andthe backgrounds, which employs the response maps gener-ated by the teacher model to instruct the student to focus ondistinguishing targets from hard negative samples. As shown in Fig. 2, each of our proposed modules con- tinuously reduces the performance gap between the student model and the teacher model without increasing the numberof parameters obviously. To sum up, our work improves an RGB-T tracker dramatically because of the following two contributions: • A Cross-Modality Distillation (CMD) framework is presented to bridge the performance gap between acompact student model and a powerful teacher model through three stages, i.e, unimodal feature extraction, multi-modal feature fusion and target state estimation.To the best of our knowledge, we are the ﬁrst to intro- duce knowledge distillation for multi-modal tracking. • Experimental results show that our proposed approach helps a student model achieves the state-of-the-art per- formance on the challenging GTOT [ 10], RGBT234 [11] and LasHer [ 14], while reducing the number of parameters and computational complexity.
Zeng_Deep_Fair_Clustering_via_Maximizing_and_Minimizing_Mutual_Information_Theory_CVPR_2023	Abstract Fair clustering aims to divide data into distinct clusters while preventing sensitive attributes (e.g., gender, race, RNA sequencing technique) from dominating the clustering. Although a number of works have been conducted and achieved huge success recently, most of them are heuristi- cal, and there lacks a unified theory for algorithm design. In this work, we fill this blank by developing a mutual information theory for deep fair clustering and accordingly designing a novel algorithm, dubbed FCMI. In brief, through maximizing and minimizing mutual information, FCMI is designed to achieve four characteristics highly ex- pected by deep fair clustering, i.e., compact, balanced, and fair clusters, as well as informative features. Besides the contributions to theory and algorithm, another contribution of this work is proposing a novel fair clustering metric built upon information theory as well. Unlike existing evalua- tion metrics, our metric measures the clustering quality and fairness as a whole instead of separate manner. To verify the effectiveness of the proposed FCMI, we conduct experiments on six benchmarks including a single-cell RNA-seq atlas compared with 11 state-of-the-art methods in terms of five metrics. The code could be accessed from https://pengxi.me .	1. Introduction Clustering plays an important role in machine learning [19, 27–29, 34, 42, 43], which could partition data into dif- ferent clusters without any label information. It has been widely used in many real-world applications such as multi- view learning [35,39], image segmentation [24], and bioin- formatics [20]. In practice, however, the data might be con- founded with sensitive attributes ( e.g., gender, race, etc., also termed as group information ) that probably over- ∗Equal contribution †Corresponding author maximize minimize Cluster Information Group InformationFigure 1. Illustration of our basic idea using information diagrams, where X,G, and Cdenote the inputs, sensitive attributes, and clustering assignments, respectively. To cluster data, we expect the overlap between non-group information and cluster informa- tion ( i.e., the conditional mutual information I(X;C\|G)) to be maximized. Meanwhile, to prevent sensitive attributes from domi- nating the clustering results, we expect the overlap between group information and cluster information ( i.e., the mutual information I(G, C )) to be minimized. whelm the intrinsic semantic of samples (also termed as cluster information ). Taking single-cell RNA clustering as a showcase, standard methods would partition data based on sequencing techniques (group information) instead of intrinsic cell types (cluster information), since cells se- quenced by different techniques would result in different expression levels [36] and most clustering methods cannot distinguish these two kinds of information. The case is sim- ilar in many automatic learning systems where the clus- tering results are biased toward sensitive attributes, which would interfere with the decision-making [9, 12, 18]. No- tably, even though these sensitive attributes are known in prior, it is daunting to alleviate or even eliminate their in- fluence, e.g., removing the “gender” information from the photos of users. As a feasible solution, fair clustering aims to hide sen- sitive attributes from the clustering results. Commonly, a clustering result is considered fair when samples of different sensitive attributes are uniformly distributed in clusters so This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 23986 that the group information is protected. However, it would lead to a trivial solution if the fairness is over-emphasized, i.e., all samples are assigned to the same cluster. Hence, in addition to fairness, balance and compactness are also highly expected in fair clustering. Specifically, a balanced clustering could avoid the aforementioned trivial solution brought by over-emphasized fairness, and the compactness refers to a clear cluster boundary. To achieve fair clustering, many studies have been con- ducted to explore how to incorporate fairness into luster- ing [3,4,8,22,23,38,44,46]. Their main differences lie in i) the stage of fairness learning, and ii)the depth of the model. In brief, [3, 8] incorporate the fairness in a pre-processing fashion by packing data points into so-called fairlets with balanced demographic groups and then partitioning them with classic clustering algorithms. [22,46] are in-processing methods that formulate fairness as a constraint for cluster- ing. As a representative of post-processing methods, [4] first performs classic clustering and then transforms the clustering result into a fair one by linear programming. Dif- ferent from the above shallow models, [23, 38, 44] propose performing fair clustering in the latent space learned by different deep neural networks to boost performance. Al- though promising results have been achieved by these meth- ods, almost all of them are heuristically and empirically de- signed, with few theoretical explanations and supports. In other words, it still lacks a unified theory to guide the algo- rithm design. In this work, we unify the deep fair clustering task un- der the mutual information theory and propose a novel theoretical-grounded deep fair clustering method accord- ingly. As illustrated in Fig. 1, we theoretically show that clustering could be achieved by maximizing the condi- tional mutual information (CMI) I(X;C\|G)between in- putsXand cluster assignments Cgiven sensitive attributes G. Meanwhile, we prove that the fairness learning could be formulated as the minimization of the mutual informa- tion (MI) I(G;C). In this case, sensitive attributes will be hidden in the cluster assignments and thus fair clustering could be achieved. To generalize our theory to deep neu- ral networks, we additionally show a deep variant could be developed by maximizing the mutual information I(X;X′) between the input Xand its approximate posterior X′. No- tably, some deep clustering methods [17,26] have been pro- posed based on the information theory. However, they are remarkably different from this work. To be exact, they ig- nored the group information. As a result, the group infor- mation will leak into cluster assignments, leading to unfair partitions. In addition, we prove that our mutual informa- tion objectives intrinsically correspond to four characteris- tics highly expected in deep fair clustering, namely, com- pact, balanced, and fair clusters, as well as informative fea- tures.Besides the above contributions to theory and algorithm, this work also contributes to the performance evaluation. To be specific, we notice that almost all existing methods evaluate clustering quality and fairness separately. How- ever, as fair clustering methods usually make a trade-off be- tween these two aspects, such an evaluation protocol might be partial and inaccurate. As an improvement, we design a new evaluation metric based on the information theory, which simultaneously measures the clustering quality and fairness. The contribution of this work could be summa- rized as follows: • We formulate deep fair clustering as a unified mu- tual information optimization problem. Specifically, we theoretically show that fair clustering could be achieved by maximizing CMI between inputs and clus- ter assignments given sensitive attributes while mini- mizing MI between sensitive attributes and cluster as- signments. Moreover, the informative feature extrac- tion could be achieved by maximizing MI between the input and its approximate posterior. • Driven by our unified mutual information theory, we propose a deep fair clustering method and carry out ex- tensive experiments to show its superiority on six fair clustering benchmarks, including a single-cell RNA at- las. • To evaluate the performance of fair clustering more comprehensively, we design a novel metric that mea- sures the clustering quality and fairness as a whole from the perspective of information theory.
Zhang_NeuralDome_A_Neural_Modeling_Pipeline_on_Multi-View_Human-Object_Interactions_CVPR_2023	Abstract Humans constantly interact with objects in daily life tasks. Capturing such processes and subsequently conduct- ing visual inferences from a fixed viewpoint suffers from occlusions, shape and texture ambiguities, motions, etc. To mitigate the problem, it is essential to build a train- ing dataset that captures free-viewpoint interactions. We construct a dense multi-view dome to acquire a complex human object interaction dataset, named HODome, that consists of ∼71M frames on 10 subjects interacting with 23 objects. To process the HODome dataset, we develop NeuralDome, a layer-wise neural processing pipeline tai- * These authors contributed equally. †Corresponding author.lored for multi-view video inputs to conduct accurate track- ing, geometry reconstruction and free-view rendering, for both human subjects and objects. Extensive experiments on the HODome dataset demonstrate the effectiveness of NeuralDome on a variety of inference, modeling, and ren- dering tasks. Both the dataset and the NeuralDome tools will be disseminated to the community for further devel- opment, which can be found at https://juzezhang. github.io/NeuralDome	1. Introduction A key task of computer vision is to understand how hu- mans interact with the surrounding world, by faithfully cap- turing and subsequently reproducing the process via mod- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 8834 eling and rendering. Successful solutions benefit broad ap- plications ranging from sports training to vocational educa- tion, digital entertainment to tele-medicine. Early solutions [11, 12, 51] that reconstruct dynamic meshes with per-frame texture maps are time-consuming and vulnerable to occlusions or lack of textures. Recent advances in neural rendering [37, 62, 71] bring huge po- tential for human-centric modeling. Most notably, the variants of Neural Radiance Field (NeRF) [37] achieve compelling novel view synthesis, which can enable real- time rendering performance [38, 59, 67] even for dynamic scenes [45, 63, 77], and can be extended to the generative setting without per-scene training [23, 69, 80]. However, less attention is paid to the rich and diverse interactions be- tween humans and objects, mainly due to the severe lack of dense-view human-object datasets. Actually, existing datasets of human-object interactions are mostly based on optical markers [61] or sparse RGB/RGBD sensors [6, 26], without sufficient appearance supervision for neural render- ing tasks. As a result, the literature on neural human-object rendering [21,57] is surprisingly sparse, let alone further ex- ploring the real-time or generative directions. Besides, ex- isting neural techniques [52,77] suffer from tedious training procedures due to the human-object occlusion, and hence infeasible for building a large-scale dataset. In a nutshell, despite the recent tremendous thriving of neural rendering, the lack of both a rich dataset and an efficient reconstruction scheme constitute barriers in human-object modeling. In this paper, we present NeuralDome , a neural pipeline that takes multi-view dome capture as inputs and conducts accurate 3D modeling and photo-realistic rendering of com- plex human-object interaction. As shown in Fig. 1, Neural- Dome exploits layer-wise neural modeling to produce rich and multi-modality outputs including the geometry of dy- namic human, object shapes and tracked poses, as well as a free-view rendering of the sequence. Specifically, we first capture a novel human-object dome (HODome) dataset that consists of 274 human-object inter- acting sequences, covering 23 diverse 3D objects and 10 human subjects (5 males and 5 females) in various appar- els. We record multi-view video sequences of natural inter- actions between the human subjects and the objects where each sequence is about 60s in length using a dome with 76 RGB cameras, resulting in 71 million video frames. We also provide an accurate pre-scanned 3D template for each object and utilize sparse optical markers to track individual objects throughout the sequences. To process the HODome dataset, we adopt an extended Neural Radiance Field (NeRF) pipeline. The brute-force adoption of off-the-shelf neural techniques such as Instant- NSR [79] and Instant-NGP [38], although effective, do not separate objects from human subjects and therefore lack sufficient fidelity to model their interactions. We insteadintroduce a layer-wise neural processing pipeline. Specif- ically, we first perform a joint optimization based on the dense inputs for accurately tracking human motions using the parametric SMPL-X model [43] as well as localizing objects using template meshes. We then propose an ef- ficient layer-wise neural rendering scheme where the hu- mans and objects are formulated as a pose-embedded dy- namic NeRF and a static NeRF with tracked 6-DoF rigid poses, respectively. Such a layer-wise representation ef- fectively exploits temporal information and robustly tack- les the occluded regions under interactions. We further in- troduce an object-aware ray sampling strategy to mitigate artifacts during layer-wise training, as well as template- aware geometry regularizers to enforce contact-aware de- formations. Through weak segmentation supervision, we obtain the decoupled and occlusion-free appearances for both the humans and the objects at a high fidelity amenable for training the input multi-view inputs for a variety of tasks from monocular motion capture to free-view rendering from sparse multi-view inputs. To summarize, our main contributions include: • We introduce NeuralDome, a neural pipeline, to ac- curately track humans and objects, conduct layer-wise geometry reconstruction, and enable novel-view syn- thesis, from multi-view HOI video inputs. • We collect a comprehensive dataset that we call HODome that will be disseminated to the community, with both raw data and the output modalities including separated geometry and rendering of individual objects and human subjects, their tracking results, free-view rendering results, etc. • We demonstrate using the dataset to train networks for a variety of visual inference tasks with complex human object interactions.
Zhang_Blind_Image_Quality_Assessment_via_Vision-Language_Correspondence_A_Multitask_Learning_CVPR_2023	Abstract We aim at advancing blind image quality assessment (BIQA), which predicts the human perception of image quality without any reference information. We develop a general and automated multitask learning scheme for BIQA to exploit auxiliary knowledge from other tasks, in a way that the model parameter sharing and the loss weighting are determined automatically. Specifically, we first describe all candidate label combinations (from multiple tasks) us- ing a textual template, and compute the joint probability from the cosine similarities of the visual-textual embed- dings. Predictions of each task can be inferred from the joint distribution, and optimized by carefully designed loss functions. Through comprehensive experiments on learn- ing three tasks - BIQA, scene classification, and distor- tion type identification, we verify that the proposed BIQA method 1) benefits from the scene classification and dis- tortion type identification tasks and outperforms the state- of-the-art on multiple IQA datasets, 2) is more robust in the group maximum differentiation competition, and 3) re- aligns the quality annotations from different IQA datasets more effectively. The source code is available at https: //github.com/zwx8981/LIQE .	1. Introduction As a fundamental computational vision task, blind im- age quality assessment (BIQA) [63] aims to predict the visual quality of a digital image with no access to the underlying pristine-quality counterpart (if any). In the age of deep learning, the development of BIQA can be mainly characterized by strategies to mitigate the conflict between the large number of trainable parameters and the *Corresponding author. (a) (b) (c) Figure 1. (a)A “parrots” image of pristine quality. (b)A dis- torted version of (a) by global Gaussian blurring. (c)A distorted “cityscape” image by the same level of Gaussian blurring. Humans are able to “see through” the Gaussian blur, and recognize the two parrots in (b) with no effort, suggesting the internal representations for the task of visual recognition should be distortion-insensitive . This makes it conceptually conflicting to BIQA, which relies on distortion-sensitive representations for quality prediction. small number of human quality annotations in the form of mean opinion scores (MOSs). When synthetic distor- tions ( e.g., Gaussian noise and JPEG compression) are of primary concern, patchwise training [4], quality-aware pre-training [32, 37, 76], and learning from noisy pseudo- labels [2, 38, 67] are practical training tricks with less (or no) reliance on MOSs. Here the underlying assumptions are that 1) the pristine-quality images exist and are acces- sible, 2) the visual distortions can be simulated efficiently and automatically, and 3) full-reference IQA models [64] are applicable and provide adequate quality approxima- tions. However, all these assumptions do not hold when it comes to realistic camera distortion ( e.g., sensor noise, motion blurring or a combination of both). A different set of training tricks have been explored, including trans- fer learning [17, 76], meta learning [80], and contrastive learning [40]. Emerging techniques that combine multiple datasets for joint training [78] and that identify informative samples for active fine-tuning [66] can also be seen as ways to confront the data challenge in BIQA. In this paper, we aim to accomplish something in the This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 14071 same spirit, but from a different multitask learning perspec- tive. We ask the key question: Can BIQA benefit from auxiliary knowledge pro- vided by other tasks in a multitask learning setting? This question is of particular interest because many high- level computer vision tasks ( e.g., object recognition [8] and scene classification [5]), with easier-to-obtain ground-truth labels, seem to be conceptually conflicting to BIQA. This is clearly illustrated in Fig. 1. Humans are able to “see through” the Gaussian blur, and recognize effortlessly the two parrots in (b). That is, if we would like to develop computational methods for the same purpose, they should rely on distortion-insensitive features, and thus be robust to such corruptions. This is also manifested by the com- mon practice in visual recognition that treats synthetic dis- tortions as forms of data augmentation [16]. In stark con- trast, BIQA relies preferentially on distortion-sensitive fea- tures to quantify the perceptual quality of images of vari- ous semantic content. Ma et al. [37] proposed a cascaded multitask learning scheme for BIQA, but did not investigate the relationships between BIQA and high-level vision tasks. Fang et al. [9] included scene classification as one task, but required manually specifying the parameters ( i.e., computa- tions) to share across tasks, which is difficult and bound to be suboptimal. Taking inspiration from recent work on vision-language pre-training [49], we propose a general and automated mul- titask learning scheme for BIQA, with an attempt to answer the above-highlighted question. Here, “automated” means that the model parameter sharing for all tasks and the loss weighting assigned to each task are determined automati- cally. We consider two additional tasks, scene classification and distortion type identification, the former of which is conceptually conflicting to BIQA, while the latter is closely related. We first summarize the scene category, distortion type, and quality level of an input image using a textual template. For example, Fig. 1 (c) may be described as “a photo of a cityscape with Gaussian blur artifacts, which is ofbadquality.” We then employ the contrastive language- image pre-training (CLIP) [49], a joint vision and language model trained with massive image-text pairs, to obtain the visual and textual embeddings. The joint probability over the three tasks can be computed from the cosine similarities between the image embedding and all candidate textual em- beddings1. We marginalize the joint distribution to obtain the marginal probability for each task, and further convert the discretized quality levels to a continuous quality score using the marginal distribution as the weighting. We supplement existing IQA datasets [7, 11, 17, 22, 27, 54] with scene category and distortion type labels, and 1We specify nine scene categories, eleven distortion types, and five quality levels, giving rise to 495textual descriptions/embeddings in total.jointly optimize the entire method on a combination of them by minimizing a weighted sum of three fidelity losses [58], where the loss weightings are adjusted automatically based on the training dynamics [31]. From extensive experimen- tal results, we arrive at a positive answer to the highlighted question: BIQA can indeed benefit from both scene clas- sification and distortion type identification. The resulting model, which we name Language- Image Quality Evaluator (LIQE), not only outperforms state-of-the-art BIQA meth- ods [13,55,76,78] in terms of prediction accuracy on multi- ple IQA datasets, but also exhibits improved generalizabil- ity in the group maximum differentiation (gMAD) competi- tion [35]. In addition, we provide quantitative evidence that LIQE better realigns MOSs from different IQA datasets in a common perceptual scale [48].
Yu_MAGVIT_Masked_Generative_Video_Transformer_CVPR_2023	Abstract We introduce the MAsked Generative VIdeo Transformer, MAGVIT , to tackle various video synthesis tasks with a single model. We introduce a 3D tokenizer to quantize a video into spatial-temporal visual tokens and propose an embedding method for masked video token modeling to fa- cilitate multi-task learning. We conduct extensive experi- ments to demonstrate the quality, efficiency, and flexibility of MAGVIT. Our experiments show that (i) MAGVIT per- forms favorably against state-of-the-art approaches and es- tablishes the best-published FVD on three video generation benchmarks, including the challenging Kinetics-600. (ii) MAGVIT outperforms existing methods in inference time by two orders of magnitude against diffusion models and by 60×against autoregressive models. (iii) A single MAGVIT model supports ten diverse generation tasks and general- izes across videos from different visual domains. The source code and trained models will be released to the public at https://magvit.cs.cmu.edu . ∗Work partially done during a research internship at Google Research.	1. Introduction Recent years have witnessed significant advances in image and video content creation based on learning frameworks ranging from generative adversarial networks (GANs) [15, 43, 48, 59, 66], diffusion models [25, 33, 35, 47, 65], to vision transformers [44, 45, 69]. Inspired by the recent success of generative image transformers such as DALL·E [46] and other approaches [12,18,20,73], we pro- pose an efficient and effective video generation model by leveraging masked token modeling and multi-task learning. We introduce the MAsked Generative VIdeo Trans- former ( MAGVIT ) for multi-task video generation. Specif- ically, we build and train a single MAGVIT model to per- form a variety of diverse video generation tasks and demon- strate the model’s efficiency, effectiveness, and flexibility against state-of-the-art approaches. Fig. 1(a) shows the quality metrics of MAGVIT on a few benchmarks with ef- ficiency comparisons in (b), and generated examples under different task setups such as frame prediction/interpolation, out/in-painting, and class conditional generation in (c). This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 10459 MAGVIT models a video as a sequence of visual tokens in the latent space and learns to predict masked tokens with BERT [17]. There are two main modules in the proposed framework. First, we design a 3D quantization model to tokenize a video, with high fidelity, into a low-dimensional spatial-temporal manifold [21, 71]. Second, we propose an effective masked token modeling (MTM) scheme for multi- task video generation. Unlike conventional MTM in image understanding [67] or image/video synthesis [12,26,28], we present an embedding method to model a video condition using a multivariate mask and show its efficacy in training. We conduct extensive experiments to demonstrate the quality, efficiency, and flexibility of MAGVIT against state- of-the-art approaches. Specifically, we show that MAGVIT performs favorably on two video generation tasks across three benchmark datasets, including UCF-101 [55], BAIR Robot Pushing [19, 61], and Kinetics-600 [10]. For the class-conditional generation task on UCF-101, MAGVIT reduces state-of-the-art FVD [61] from 332 [21] to 76 (↓77%). For the frame prediction task, MAGVIT performs best in terms of FVD on BAIR ( 84[35]→62,↓26%) and Kinetics-600 ( 16[33]→9.9,↓38%) . Aside from the visual quality, MAGVIT’s video synthe- sis is highly efficient. For instance, MAGVIT generates a 16-frame 128 ×128 video clip in 12steps, which takes 0.25 seconds on a single TPUv4i [36] device. On a V100 GPU, a base variant of MAGVIT runs at 37frame-per-second (fps) at 128×128 resolution. When compared at the same res- olution, MAGVIT is two orders of magnitude faster than the video diffusion model [33]. In addition, MAGVIT is 60
Zhang_Coaching_a_Teachable_Student_CVPR_2023	Abstract We propose a novel knowledge distillation framework for effectively teaching a sensorimotor student agent to drive from the supervision of a privileged teacher agent. Cur- rent distillation for sensorimotor agents methods tend to re- sult in suboptimal learned driving behavior by the student, which we hypothesize is due to inherent differences between the input, modeling capacity, and optimization processes of the two agents. We develop a novel distillation scheme that can address these limitations and close the gap between the sensorimotor agent and its privileged teacher. Our key in- sight is to design a student which learns to align their input features with the teacher’s privileged Bird’s Eye View (BEV) space. The student then can benefit from direct supervision by the teacher over the internal representation learning. To scaffold the difficult sensorimotor learning task, the student model is optimized via a student-paced coaching mecha- nism with various auxiliary supervision. We further propose a high-capacity imitation learned privileged agent that sur- passes prior privileged agents in CARLA and ensures the student learns safe driving behavior. Our proposed sen- sorimotor agent results in a robust image-based behavior cloning agent in CARLA, improving over current models by over 20.6%in driving score without requiring LiDAR, historical observations, ensemble of models, on-policy data aggregation or reinforcement learning.	1. Introduction Learning internal representations for making intricate driving decisions from images involves a complex optimiza- tion task [4, 43, 54]. The inherent challenge for end-to-end training of driving agents lies in the immense complexity of learning to map high-dimensional visual observations into general and safe navigational decisions [13, 20, 70]. Even given millions of training examples [4], today’s agents still fail to reliably learn an internal representation that can be used for robust processing of complex visual scenarios (e.g., dense urban settings with intricate layouts and dynamic ob- stacles) in a safe and task-driven manner [13, 20, 70]. To ease the challenging sensorimotor training task, re- Privileged Agent Teachable NetBEV Representation RGB Image StudentTeacher Bicyclist Alignment ModuleCoachFigure 1. Effective Knowledge Distillation for Sensorimotor Agents. Our proposed CaT (Coaching a Teachable student) frame- work enables highly effective knowledge transfer between a priv- ileged teacher and a sensorimotor (i.e., image-based) student. An alignment module learns to transform image-based features to the teacher’s BEV feature space, where the student can then lever- age extensive and direct supervision on its learned intermediate representations. The student model is optimized via a coaching mechanism with extensive auxiliary supervision in order to fur- ther scaffold the difficult sensorimotor learning task. cent approaches decompose the task into stages, e.g., by first training a high-capacity privileged network with com- plete knowledge of the world and distilling its knowl- edge into a less capable vision-based student network [11, 13, 23, 47, 79]. However, due to the inherent differences between the inputs and architectures of the two agents, current methods rely on limited supervisory mechanisms from the teacher, i.e., exclusively through the teacher’s output [11, 13] or knowledge distillation of a single final fully-connected layer [71, 79, 81]. Moreover, the privileged teacher’s demonstration targets may be noisy or difficult for the student to imitate, given the limited perspective [27]. In this work, we sought to develop a more effective knowledge distillation paradigm for training a sensorimotor agent to drive. Our key insight is to enable more extensive supervi- sion from the teacher by reducing the gap between internal modeling and learning capabilities between the two agents. Our proposed approach for holistic knowledge distilla- tion is informed by human instruction, which often involves structured supervision in addition to high-level demonstra- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 7805 tions, e.g., providing various hints to scaffold information in a way that the student can better understand [30]. When teaching others new and challenging skills, i.e., where a stu- dent may not be able to replicate the demonstration such as riding a bicycle or driving a vehicle, teachers may provide additional supervision regarding the underlying task struc- ture and their own internal reasoning [64]. Analogously to our settings, the privileged teaching agent can potentially provide richer supervision when teaching a limited capacity sensorimotor student, i.e., through more careful and direct guidance of the underlying representation learning. In our work, we introduce CaT, a novel method for teach- ing a sensorimotor student to drive using supervision from a privileged teacher. Our key insights are threefold: 1) Effec- tive Teacher: We propose to incorporate explicit safety- aware cues into the BEV space that facilitate a surpris- ingly effective teacher agent design. While prior privileged agents struggle to learn to drive in complex urban driving scenes, we demonstrate our learned agent to match expert- level decision-making. 2) Teachable Student via Align- ment: An IPM-based transformer alignment module can facilitate direct distillation of most of the teacher’s features and better guide the student learning process. 3) Student- paced Coaching: A coaching mechanism for managing difficult samples can scaffold knowledge and lead to im- proved model optimization by better considering the ability of the student. By holistically tackling the complex knowl- edge distillation task with extensive teacher and auxiliary supervision, we are able to train a state-of-the-art image- based agent in CARLA [21]. Through ablation studies, in- put design, and interpretable feature analysis, we also pro- vide critical insights into current limitations in learning ro- bust and generalized representations for driving agents.
Zancato_TrainTest-Time_Adaptation_With_Retrieval_CVPR_2023	Abstract We introduce Train/Test-Time Adaptation with Re- trieval ( T3AR), a method to adapt models both at train and test time by means of a retrieval module and a searchable pool of external samples. Before inference, T3ARadapts a given model to the downstream task using reﬁned pseudo- labels and a self-supervised contrastive objective function whose noise distribution leverages retrieved real samples to improve feature adaptation on the target data manifold. The retrieval of real images is key to T3ARsince it does not rely solely on synthetic data augmentations to com- pensate for the lack of adaptation data, as typically done by other adaptation algorithms. Furthermore, thanks to the retrieval module, our method gives the user or service provider the possibility to improve model adaptation on the downstream task by incorporating further relevant data or to fully remove samples that may no longer be available due to changes in user preference after deployment. First, we show that T3ARcan be used at training time to im- prove downstream ﬁne-grained classiﬁcation over standard ﬁne-tuning baselines, and the fewer the adaptation data the higher the relative improvement (up to 13%). Second, we apply T3ARfor test-time adaptation and show that exploit- ing a pool of external images at test-time leads to more ro- bust representations over existing methods on DomainNet- 126 and VISDA-C, especially when few adaptation data are available (up to 8%).	1. Introduction While Deep Learning models are evolving rapidly, ma- chine learning systems used in production are updated rarely, as each deployment requires the provider to engage in a complex process of scaling, securitization, certiﬁcation of new model and dataset cards, bias evaluation, and re- *Work done during an internship at AWS AI Labs.Decision boundary after adaptationPre-trained decision boundaryTest samples from TSource data from S Retrieved samples AExternal datasetTTarget datasetSSSource datasetFigure 1. Adaptation with retrieval from an external data pool. Illustration of how T3ARexploits target data Tand the external data pool Ato adapt the decision boundary after pre-training on the source datasets S. For new test queries from the target dataset, T3ARapproximates the local data manifold around Tby retriev- ing similar unlabelled examples from A. Then, it updates the de- cision boundary with a contrastive self-supervised objective. gression tests. It is now common for users to adapt trained models to their speciﬁc use cases, or to the changed con- text as time goes by [ 16,39,60]. Such adaptation can be performed by ﬁne-tuning on a speciﬁc dataset Sowned by the user [ 1,18]. However, on an even ﬁner time-scale, users may want to adapt their models based on data they observe at test time, bypassing the time-consuming annotation pro- cess [ 8,35,53,57]. Test-Time Adaptation (TTA) refers to the problem of adapting a source model to a target task T represented by test data, for which no ground-truth labels are given. This trend is exacerbated by the advent of Foundation Models [ 2,10,38,59], at least in the visual domain where tasks can be antagonistic and models are sensitive to even subtle changes in the data distribution. At the same time, both users and providers typically have access to ever- growing pools of auxiliary data, albeit often heterogeneous (pertaining to concepts other than the one of interest at test- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 15911 time), and without annotations. Yet it seems plausible that, somewhere within these large pools of data, there may be information useful for the task at hand. In this paper, we tackle the problem of performing test- time adaptation by retrieving information from a large, un- labeled, heterogeneous, and evolving dataset. The same procedure could also be followed by the provider, if they have access to auxiliary internal data and wish to adapt the production model based on trends observed in test data. We refer to our method as Train/Test-Time Adaptation with Re- trieval , orT3AR. T3AR, if solved, would enable a number of real-world tasks that have thus far frustrated practitioners. For in- stance, it would allow a user to select, among a vast data lake A, which samples to use for a training, based on la- beled and unlabeled samples [ 61]. It would also enable nim- ble inference, by adapting a modest-size model to speciﬁc tasks, rather than relying on an unwieldy model to master all trades. Finally, it would enable reversible adaptation: While in the case of language models tasks are generally synergistic [ 44], in vision tasks can be antagonistic.1There- fore, a model adapted to one task may behave poorly on another, and a model that encompasses both would require signiﬁcantly higher capacity [ 2,10,38,59], to the detriment of inference efﬁciency. In T3AR, changing the target data Tchanges the subset of the data pool Athat is retrieved, with no impact on other models, instantiating smaller inde- pendent models for antagonistic tasks, rather than coercing them into a larger one, likely multiplying inference costs. T3ARcan be used in a continual setting, where at each time tone has a different target Tt, and the auxiliary task A is composed of the union of all prior targets T0,...,T t. The retrieval system should automatically determine what infor- mation from whatever past targets is relevant to the present, and what information is redundant in Aand can be elimi- nated. The important difference compared to ordinary con- tinual learning is that each step starts with the base model, so there is no catastrophic forgetting, and what is updated is the auxiliary task. In other words, the integration of infor- mation occurs in A, not in the trained model f. 1.1. Related problems T3ARrelates to unsupervised domain adaptation (UDA) [26,28,45], since the target dataset is not annotated. How- ever, in UDA one assumes that the source dataset Sis avail- able along with the target T, which is not necessarily the case in T3ARsince users may want to bypass annotation altogether, and directly adapt the pre-trained model using the auxiliary dataset A, based on the target task T, without having direct access to S. 1E.g., localization requires marginalizing identity, whereas recognition requires marginalizing location, making the features that are informative for one detrimental to the other [ 2,38].T3ARalso relates to semi-supervised learning (SSL) [32,34,51], since the target dataset Tand the auxiliary dataset Aare not annotated. However, in SSL one assumes that labeled Sand unlabeled data are drawn from the same joint distribution, which is not the case for TandAin T3AR, and, in any case we do not aim to infer labels of A, and just use it to improve the model on the target task. T3ARis also related to open-set domain adaptation [ 6, 49], since the auxiliary dataset Ais heterogeneous and does not share the same label space as the source and target task. It is also related to out-of-distribution detection (OOD) [ 20, 62], since one needs to decide whether to add samples from the auxiliary dataset, and to active learning [ 50], since one needs to decide what samples to add. Naturally, T3ARclosely relates to test-time adaptation (TTA) [ 8,35,53,57,65], and to memory-augmented or retrieval-based architectures [ 3,11,36], widely developed in the language domain [ 4,33,63], where the hypotheses live in the same space of the data and nuisance variability is limited to paraphrasing. In summary, T3ARlies at the intersection of UDA, SSL, OOD, TTA, Active Learning, and Retrieval, yet it does not ﬁt neatly into any of them, making both the survey of related literature (Sect. 2) and experimental assessment (Sect. 4) non-straightforward. 1.2. Key ideas and contributions We propose a method to solve T3AR, based on a target unlabeled dataset T, that selects samples from an auxiliary dataset A, using a retrieval model R. Starting from any model fSpre-trained by the provider on a dataset Dand later ﬁne-tuned by the user on a labelled dataset S, our method ﬁnds subsets of an auxiliary dataset Athat are relevant for the target dataset T, using nearest neighbors in Ato samples in T, measured in a representa- tion space computed by a retrieval model R(in our case, a CLIP embedding [ 48]). The key technical contribution is a contrastive loss used for updating the model fSto a new model fA\|T, whereby negative pairs are selected by retrieving samples from the external dataset Athat are informative ofTusing the re- triever R. Furthermore, to improve training stability, we exclude same-class negatives pairs from Tby exploiting as- signed pseudo-labels obtained by averaging predicted log- its on different data augmentations. Our method can be thought of as a form of contrastive “dataset augmentation” by enlarging the user data with samples drawn from a dif- ferent (unlabeled) dataset A, based on guidance provided by a retriever R. This procedure can be followed by both the user and the provider, thus empowering them to adapt the core model (train-time adaptation) or a sequence of disjoint custom models (test-time adaptation). We show that applying T3ARimproves downstream 15912 classiﬁcation accuracy over the paragon supervised ﬁne- tuning [ 1,18] for train-time and test-time adaptation meth- ods [ 8,35,57] for test-time. In particular, as the number of data available during adaptation decreases, T3ARimproves by up to 13% and 8% in relative Top1 accuracy at train and test time, respectively.
Xu_PIDNet_A_Real-Time_Semantic_Segmentation_Network_Inspired_by_PID_Controllers_CVPR_2023	Abstract Two-branch network architecture has shown its effi- ciency and effectiveness in real-time semantic segmentation tasks. However, direct fusion of high-resolution details and low-frequency context has the drawback of detailed features being easily overwhelmed by surrounding contextual infor- mation. This overshoot phenomenon limits the improvement of the segmentation accuracy of existing two-branch mod- els. In this paper, we make a connection between Convolu- tional Neural Networks (CNN) and Proportional-Integral- Derivative (PID) controllers and reveal that a two-branch network is equivalent to a Proportional-Integral (PI) con- troller, which inherently suffers from similar overshoot is- sues. To alleviate this problem, we propose a novel three- branch network architecture: PIDNet, which contains three branches to parse detailed, context and boundary informa- tion, respectively, and employs boundary attention to guide the fusion of detailed and context branches. Our family of PIDNets achieve the best trade-off between inference speed and accuracy and their accuracy surpasses all the existing models with similar inference speed on the Cityscapes and CamVid datasets. Specifically, PIDNet-S achieves 78.6% mIOU with inference speed of 93.2 FPS on Cityscapes and 80.1% mIOU with speed of 153.7 FPS on CamVid.	1. Introduction Proportional-Integral-Derivative (PID) Controller is a classic concept that has been widely applied in modern dy- namic systems and processes such as robotic manipulation [3], chemical processes [24], and power systems [25]. Even though many advanced control strategies with better con- trol performance have been developed in recent years, PID controller is still the go-to choice for most industry applica- tions due to its simplicity and robustness. Furthermore, the idea of PID controller has been extended to many other ar- 1Work supported in part by NSF grants ECCS-1923803 and CCF- 2007527. Figure 1. The trade-off between inference speed and accuracy (re- ported) for real-time models on the Cityscapes [12] test set. Blue stars refer to our models while green triangles represent others. eas. For example, researchers introduced the PID concept to image denoising [32], stochastic gradient decent [1] and nu- merical optimization [50] for better algorithm performance. In this paper, we devise a novel architecture for real-time se- mantic segmentation tasks by employing the basic concept of PID controller and demonstrate that the performance of our model surpasses all the previous works and achieves the best trade-off between inference speed and accuracy, as il- lustrated in Figure 1, by extensive experiments. Semantic segmentation is a fundamental task for visual scene parsing with the objective of assigning each pixel in the input image to a specific class label. With the increas- ing demand of intelligence, semantic segmentation has be- come the basic perception component for applications such as autonomous driving [16], medical imaging diagnosis [2] and remote sensing imagery [54]. Starting from FCN [31], which achieved great improvement over traditional meth- ods, deep convnets gradually dominated the semantic seg- mentation field and many representative models have been proposed [4, 6, 40, 48, 59, 60]. For better performance, var- ious strategies were introduced to equip these models with the capability of learning contextual dependencies among This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 19529 pixels in large scale without missing important details. Even though these models achieve encouraging segmentation ac- curacy, too much computational cost are required, which significantly hinder their application in real-time scenarios, such as autonomous vehicle [16] and robot surgery [44]. To meet real-time or mobile requirements, researchers have come up with many efficient and effective models in the past for semantic segmentation. Specifically, ENet [36] adopted lightweight decoder and downsampled the feature maps in early stages. ICNet [58] encoded small-size in- puts in complex and deep path to parse the high-level se- mantics. MobileNets [21, 42] replaced traditional convolu- tions with depth-wise separable convolutions. These early works reduced the latency and memory usage of segmen- tation models, but low accuracy significantly limits their real-world application. Recently, many novel and promis- ing models based on Two-Branch Network (TBN) architec- ture have been proposed in the literature and achieve SOTA trade-off between speed and accuracy [15, 20, 38, 39, 52]. In this paper, we view the architecture of TBNs from the prospective of PID controller and point out that a TBN is equivalent to a PI controller, which suffers from the over- shoot issue as illustrated in Figure 2. To alleviate this problem, we devise a novel three-branch network architec- ture, namely PIDNet, and demonstrate its superiority on Cityscapes [12], CamVid [5] and PASCAL Context [33] datasets. We also provide ablation study and feature visual- ization for better understanding of the functionality of each module in PIDNet. The source code can be accessed via: https://github.com/XuJiacong/PIDNet The main contributions of this paper are three-fold: • We make a connection between deep CNN and PID controller and propose a family of three-branch net- works based on the PID controller architecture. • Efficient modules, such as Bag fusion module de- signed to balance detailed and context features, are proposed to boost the performance of PIDNets. • PIDNet achieves the best trade-off between inference speed and accuracy among all the existing models. In particular, PIDNet-S achieves 78.6%mIOU with speed of 93.2FPS and PIDNet-L presents the high- est accuracy ( 80.6%mIOU) in real-time doman on Cityscapes test set without acceleration tools.
Yang_Relational_Space-Time_Query_in_Long-Form_Videos_CVPR_2023	Abstract Egocentric videos are often available in the form of un- interrupted, uncurated long videos capturing the camera wearers’ daily life activities.Understanding these videos re- quires models to be able to reason about activities, objects, and their interactions. However, current video benchmarks study these problems independently and under short, cu- rated clips. In contrast, real-world applications, e.g. AR assistants, require bundling these problems for both model development and evaluation. In this paper, we propose to study these problems in a joint framework for long video understanding. Our contributions are three-fold. First, we propose an integrated framework, namely Relational Space- Time Query (ReST), for evaluating video under- standing models via templated spatiotemporal queries. Sec- ond, we introduce two new benchmarks, ReST-ADL and ReST-Ego4D1, which augment the existing egocentric video datasets with abundant query annotations generated by the ReST framework. Finally, we present a set of baselines and 1The latest version of our benchmark and models will be available here.in-depth analysis on the two benchmarks and provide in- sights about the query tasks. We view our integrated frame- work and benchmarks as a step towards comprehensive, multi-step reasoning in long videos, and believe it will fa- cilitate the development of next generations of video under- standing models.	1. Introduction Thanks to the advances of modern, massive parallel hardware, e.g. GPUs, and the availability of large datasets, significant progress has been made in the last few years with large language models ( e.g., GPT-3 [6], BERT [11]) and image / video generative models ( e.g., DALLE [40], Imagen [41], Make-A-Video [46]). Meanwhile, current video understanding models mostly focus on processing short video clips [12, 13, 52] and solving basic perception tasks such as action recognition [16, 28, 31, 44, 48] and de- tection [7, 19]. One may ask the questions for video under- standing research: “How far are current models progressing to a human-level performance on video understanding?”, or This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 6398 “What is blocking us from building models that can under- stand complex relationships in long videos?” Of course, there exists multiple blockers in practice such as GPU memory limitation and inefficient hardware support for processing long videos. Yet the first and most important reason is always the lack of the right research problem and theright benchmark . One drawback of current video un- derstanding benchmarks [19, 28] is that they handle analy- sis of activities, objects and their interactions in a separate manner. However, understanding long-form videos usually requires a unified analysis of these factors because activi- ties manifesting within these uncurated videos are primar- ily in the form of human-object interaction, especially for egocentric recordings of a camera wearer’s daily lives [17]. In recent years, video-QA [18, 33, 49, 66] and video cap- tioning [10, 50] have been proposed as alternative tasks for video understanding. These tasks require models to under- stand both visual and text modalities and perform cross- modal reasoning. On one hand, such vision-language based tasks have the benefit of bypassing the pre-defined taxon- omy and closed-world assumptions by leveraging language as input and/or output. On the other hand, using language for vision tasks, either in the form of input query or output prediction, brings additional ambiguity in text generation and requires use of uninterpretable evaluation metrics ( e.g., BLEU, ROUGE). Language priors also introduce bias to the task as observed in prior work that the language-only model achieves comparable results with the VQA ones [4, 18]. In this paper, we present a holistic framework, Relational Space- Time Query (ReST), for evaluating video under- standing models via templated spatiotemporal queries. By combining analysis of activities, objects, and their interac- tions, our ReST framework captures much of the rich ex- pressivity of natural language query while avoiding the am- biguity and bias introduced by the language input / output. Figure 1 illustrates an example of our ReST framework. Given a long video spanning up to 30 minutes, we evalu- ate a video understanding model by asking various queries about the activities and human-object interactions occurred in the video. Unlike VQA that relies on language-based questions and answers, all of our queries and answers are constructed in the form of pre-defined templates with vi- sual or categorical input. Such a design helps the evaluation remain pure vision-focused and enjoy well-defined, well- posed evaluation metrics. Queries constructed in ReST can cover various questions with different levels of complexity. As shown in the examples in Figure 1, the questions can be: “what activities did I do with the coffee mug?”, “what objects did I interact with when I was washing dishes?” or “when did I take the pills stored in this specific bottle?”. We note that these questions are templated and only the time in square brackets, the activities in the double quotes, and the image crops in the curvy brackets are allowed to varyto form different questions. In order to perform well on these query tasks, a model needs not only be able to process the long video efficiently, but is also required to understand the temporal dependencies of activities and the fine-grained human-object interactions across the video. We summarize our contributions as follows. We present Relational Space- Time Query (ReST), a holistic frame- work and benchmarks for long video understanding with detailed-reasoning tasks. We provide annotations for our ReST queries on two existing benchmarks, ADL [37] and Ego4D [17]. We conduct a set of comprehensive baselines and analysis for ReST to gain insights into the tasks. We find that even with the initial set of the three basic tasks, current state-of-the-art models are not meeting desired per- formance, which indicates the need for further research and opportunities in this field.
Ye_Improving_Commonsense_in_Vision-Language_Models_via_Knowledge_Graph_Riddles_CVPR_2023	Abstract This paper focuses on analyzing and improving the com- monsense ability of recent popular vision-language (VL) models. Despite the great success, we observe that existing VL-models still lack commonsense knowledge/reasoning ability (e.g., “Lemons are sour”), which is a vital com- ponent towards artificial general intelligence. Through our analysis, we find one important reason is that exist- ing large-scale VL datasets do not contain much common- sense knowledge, which motivates us to improve the com- monsense of VL-models from the data perspective. Rather than collecting a new VL training dataset, we propose a more scalable strategy, i.e., “Data Augmentation with kNowledge graph linearization for CommonsensE capabil- ity” (DANCE). It can be viewed as one type of data aug- mentation technique, which can inject commonsense knowl- edge into existing VL datasets on the fly during training. More specifically, we leverage the commonsense knowl- edge graph (e.g., ConceptNet) and create variants of text description in VL datasets via bidirectional sub-graph se- quentialization. For better commonsense evaluation, we further propose the first retrieval-based commonsense di- agnostic benchmark. By conducting extensive experiments on some representative VL-models, we demonstrate that our DANCE technique is able to significantly improve the commonsense ability while maintaining the performance on vanilla retrieval tasks. The code and data are available at https://github.com/pleaseconnectwifi/DANCE.	1. Introduction Many vision-based problems in our daily life go beyond perception and recognition. For example, when we hear people say “ It tastes sour ”, we need to identify they are talk- ing about lemons on the table instead of the chocolate cake. Therefore, it is essential for artificial general intelligence to *This work was partially supported by a GRF grant (Project No. CityU 11216122) from the Research Grants Council (RGC) of Hong Kong. This tastes sour . B CLIP BLIP (itc) BLIP (itm) ViLT AFigure 1. Illustration of the commonsense lacking problem of various popular VL-models, including CLIP [45] pre-trained with contrastive supervision, ViLT [24] with matching supervision, and BLIP [28] with the both. The bar plots suggest the alignment scores of the images to the text. All models fail in retrieving the correct image with lemon (in blue). develop commonsense capability. Vision-Language (VL) models [45] recently show promising signals on mimick- ing the core cognitive activities of humans by understand- ing the visual and textual information in the same latent space [74]. However, we observed that VL-models, e.g., CLIP [45], still struggle when minor commonsense knowl- edge is needed. For example, as shown in figure 1, none of the existing models correctly identify the lemon with text input “ It tastes sour ”. In this work, we take a step towards injecting the VL- models with commonsense capability. More specifically, we find one important reason for the commonsense lacking issue is that existing large-scale VL datasets do not con- tain much commonsense knowledge. On the one hand, reg- ular VL data, e.g., COCO [32] and CC 12M [9] contain more nouns and descriptive adjectives, with much fewer verbs and particles compared to regular texts. This distribu- tion difference suggests that it might be infeasible for VL- models to gain commonsense ability by purely enlarging the dataset, unlike language-only models [22, 44]. Also, other objectives like visual question answering or generation are This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 2634 not widely applicable for training and have limited data size. Inspired by the aforementioned findings, we propose Data Augmentation with kNowledge graph linearization for CommonsensE capability (DANCE). The main idea is to generate commonsense-augmented image-text pairs. To do so, one natural idea is to leverage the rich commonsense knowledge in knowledge graphs [5, 56]. However, it is not trivial to inject the knowledge into image-text pairs. On the one hand, structured data like graphs usually require spe- cific architectures [59, 75] to embed, which is troublesome. On the other hand, if we associate the external knowledge with the text in the training stage, we will need the exter- nal knowledge-augmentation process in the inference stage as well to avoid domain shift [51]. This is not desirable, since the corresponding knowledge is usually not available for the inference tasks. To address these challenges, we first re-organize the commonsense knowledge graph into entries with (entity, relation, entity) format, and pair them to the images that contain one of the entities. We then hide the name of entities in that image with demonstrative pronouns, e.g., “ this item ”. The generated descriptions are in tex- tual form and therefore readily applicable for the training of most VL-models. More importantly, by forcing the model to memorize the relationships between entities in the train- ing stage, such data augmentation is not needed in the infer- ence stage. The data pair generation pipeline is automatic and scalable, leveraging the existing consolidated common- sense knowledge base and the large and various collections of image-language supervision. In addition, existing VL commonsense evaluations are restricted to visual question answering and generation which are not a good fit or well received in the majority of VL-models. Therefore, we propose a new diagnostic test set in a wider adaptable form, i.e., Image-Text and Text-Image Retrieval, to achieve a fair evaluation of the pre-trained VL- models. The set is upgraded by neighborhood hard-negative filtering to further ensure data quality. The effectiveness of the DANCE is validated by not only our diagnostic test set, but also the most popular vi- sual question answering benchmark for commonsense [36]. Moreover, we show the commonsense ability of the models trained with DANCE even generalize to unseen knowledge. We show the potential of the new train strategy and the test dataset by deep content analysis and baseline performance measurements across various cutting-edge VL-models. We summarize our main findings and contributions as follows: 1. We propose a novel commonsense-aware training strategy DANCE, which is compatible with the most of VL-models. The inference stage needs no change. 2. We propose a new retrieval-based well-received com- monsense benchmark to analyze a suite of VL-models and discover weaknesses that are not widely known:commonsense easy for humans (83%) is hard for cur- rent state-of-the-art VL-models ( <42%). 3. We conduct extensive experiments to demonstrate the effectiveness of the proposed strategy and diagnostic test set. The datasets and all the code will be made publicly available.
Xu_Side_Adapter_Network_for_Open-Vocabulary_Semantic_Segmentation_CVPR_2023	Abstract This paper presents a new framework for open- vocabulary semantic segmentation with the pre-trained vision-language model, named Side Adapter Network (SAN). Our approach models the semantic segmentation task as a region recognition problem. A side network is attached to a frozen CLIP model with two branches: one for predicting mask proposals, and the other for predicting attention bias which is applied in the CLIP model to rec- ognize the class of masks. This decoupled design has thebenefit CLIP in recognizing the class of mask proposals. Since the attached side network can reuse CLIP features, it can be very light. In addition, the entire network can be trained end-to-end, allowing the side network to be adapted to the frozen CLIP model, which makes the predicted mask proposals CLIP-aware. Our approach is fast, accurate, and only adds a few additional trainable parameters. We evalu- ate our approach on multiple semantic segmentation bench- marks. Our method significantly outperforms other coun- terparts, with up to 18 times fewer trainable parameters and 19 times faster inference speed. Fig. 1 shows some visual- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 2945 ization results on ImageNet. We hope our approach will serve as a solid baseline and help ease future research in open-vocabulary semantic segmentation.	1. Introduction Recognizing and segmenting the visual elements of any category is the pursuit of semantic segmentation. Mod- ern semantic segmentation methods [5, 7, 23] rely on large amounts of labeled data, but typically datasets often only consist of tens to hundreds of categories, and expensive data collection and annotation limit our possibilities to fur- ther expand the categories. Recently, large-scale vision- language models [17, 27, 36,37], represented by CLIP [27], have enabled arbitrary category recognition at the image level, i.e., open-vocabulary image classification , and this great success encourages us to explore its adaptation in se- mantic segmentation. Applying the CLIP model in open-vocabulary seman- tic segmentation is challenging because the CLIP model is trained by image-level contrastive learning. Its learned rep- resentation lacks the pixel-level recognition capability that is required for semantic segmentation. One solution [12,19] to remedy the granularity gap of representation is fine- tuning the model on the segmentation dataset. However, the data sizes of segmentation datasets are much less than the vision-language pre-training dataset, so the capability of fine-tuned models on open-vocabulary recognition is often compromised. Modeling semantic segmentation as a region recogni- tion problem bypasses the above difficulties. Early at- tempts [9, 33] adopt a two-stage training framework. In the first stage, a stand-alone model is trained to generate a set of masked image crops as mask proposals. In the second stage, the vision-language pre-training model ( e.g. CLIP) is used to recognize the class of masked image crops. How- ever, since the mask prediction model is completely inde- pendent of the vision-language pre-training model, it misses the opportunity to leverage the strong features of the vision- language pre-training model and the predicted masked im- age crops may be unsuitable for recognition, which leads to a heavy, slow, and low-performing model. This work seeks to fully unleash the capabilities of the vision-language pre-training model in open vocabulary se- mantic segmentation. To reach this goal, we present a new framework (Fig. 2), called side adapter network (SAN). Its mask prediction and recognition are CLIP-aware because of end-to-end training, and it can be lightweight due to lever- aging the features of CLIP. The side adapter network has two branches: one predict- ing mask proposals, and one predicting attention biases that are applied to the self-attention blocks of CLIP for mask class recognition. We show this decoupled design improves 𝑓𝑓 𝑓𝑓 𝑓𝑓Attention Bias Mask ProposalsProposal LogitsCLIP Model Side Adapter Network For InferencePredictionFigure 2. Overview of our SAN. The red dotted lines indicate the gradient flow during training. In our framework, the frozen CLIP model still serves as a classifier, and the side adapter network gen- erates mask proposals and attention bias to guide the deeper layers of the CLIP model to predict proposal-wise classification logits. During inference, the mask proposals and the proposal logits are combined to get final predictions through Matmul . the segmentation performance because the region used for CLIP to recognize the mask may be different from the mask region itself. To minimize the cost of CLIP, we further present a single-forward design: the features of shallow CLIP blocks are fused to SAN, and other deeper blocks are combined with attention biases for mask recognition. Since the training is end-to-end, the side adapter network can be maximally adapted to the frozen CLIP model. With the aim of fairness and reproducibility, our study is based on officially released CLIP models. We focus on the released ViT CLIP models because the vision transformer hasde facto substituted ConvNet as the dominant backbone in the computer vision community, and for conceptual con- sistency and simplicity, the side adapter network is also im- plemented by the vision transformer. Accurate semantic segmentation needs high-resolution images, but the released ViT CLIP models are designed for low-resolution images ( e.g.224×224) and directly ap- ply to high-resolution images giving a poor performance. To alleviate the conflicts in input resolutions, we use low- resolution images in the CLIP model and high-resolution images in the side adapter network. We show this asym- metric input resolution is very effective. In addition, we also explore only fine-tuning the positional embedding of the ViT model and note improvements. We evaluate our method on various benchmarks. Fol- lowing the setting of previous works [22, 33], the COCO Stuff [4] dataset is used for training, and Pascal VOC [11], Pascal Context-59 [25], Pascal Context-459 [25], ADE20K- 150 [41], and ADE20K-847 [41] are used for testing. With- out bells and whistles, we report state-of-the-art perfor- mance on all benchmarks: with the CLIP ViT-L/14 model, our method achieves 12.4 mIoU on ADE-847, 15.7 mIoU on PC-459, 32.1 mIoU on ADE-150, 57.7 mIoU on PC- 59, and 94.6 mIoU on VOC. Compared to the previous best 2946 method, our method has an average of +1.8 mIoU improve- ments on 5 datasets for ViT-B/16, and +2.3 mIoU improve- ments for ViT-L/14, respectively. By further applying en- semble trick, the average performance gap increases to +2.9 mIoU and +3.7 mIoU for ViT-B/16 and ViT-L/14. Along with the excellent performance, our approach re- quires only 8.4M trainable parameters with 64.3 GFLOPs, which is only 13% and20% of [10], 6%and less than 1% of [22], respectively.
Zeng_Real-Time_Multi-Person_Eyeblink_Detection_in_the_Wild_for_Untrimmed_Video_CVPR_2023	Abstract Real-time eyeblink detection in the wild can widely serve for fatigue detection, face anti-spoofing, emotion analysis, etc. The existing research efforts generally focus on single- person cases towards trimmed video. However, multi- person scenario within untrimmed videos is also impor- tant for practical applications, which has not been well concerned yet. To address this, we shed light on this re- search field for the first time with essential contributions on dataset, theory, and practices. In particular, a large-scale dataset termed MPEblink that involves 686 untrimmed videos with 8748 eyeblink events is proposed under multi- person conditions. The samples are captured from uncon- strained films to reveal “in the wild” characteristics. Mean- while, a real-time multi-person eyeblink detection method is also proposed. Being different from the existing counter- parts, our proposition runs in a one-stage spatio-temporal way with end-to-end learning capacity. Specifically, it si- multaneously addresses the sub-tasks of face detection, face tracking, and human instance-level eyeblink detection. This paradigm holds 2 main advantages: (1) eyeblink features can be facilitated via the face’s global context (e.g., head pose and illumination condition) with joint optimization and interaction, and (2) addressing these sub-tasks in par- allel instead of sequential manner can save time remarkably to meet the real-time running requirement. Experiments on MPEblink verify the essential challenges of real-time multi- person eyeblink detection in the wild for untrimmed video. Our method also outperforms existing approaches by large margins and with a high inference speed. †Yang Xiao is corresponding author (Yang [email protected]). 8.38s 8.54s 8.71s 8.88s9.04s 9.21s 9.21s 9.38s 9.50s 9.63s Person 1 blink Person 2 blinkPerson 1 Person 2 Timeline TimelineFigure 1. The illustration on multi-person eyeblink detection. This task aims at being aware of the presence of people and detecting their eyeblink activities at instance level.	1. Introduction Real-time eyeblink detection in the wild is a recently emerged challenging research task [19] that can widely serve for fatigue detection [2], face anti-spoofing [34], af- fective analysis [7], etc. Although remarkable progress has been made [9, 10, 19], the existing methods generally fo- cus on single-person cases within trimmed videos. Multi- person scenario within untrimmed videos has not been well concerned yet. However, detecting long-term eyeblink be- haviors at multi-instance level is more preferred for some practical application scenarios. For example, it can be used to estimate attendees’ attention level and emotional state change during social interaction [9, 33]. Thus, effective and real-time multi-person eyeblink detection in the wild for untrimmed video is indeed required. To this end, we shed the light on this research problem with essential contributions on dataset, theory, and prac- tices. First, a challenging labeled multi-person eyeblink de- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 13854 tection benchmark termed MPEblink is built under in-the- wild conditions. It consists of 686 untrimmed long videos captured from unconstrained movies to reveal the “in the wild” characteristics. The contained scenarios are realistic and diverse, such as social interactions and group activi- ties. To our knowledge, MPEblink is the first multi-person eyeblink detection dataset that focuses on in-the-wild long videos. Fig. 1 illustrates a sample video with ground truth annotations within it. Different from the existing eyeblink detection benchmarks [10,12,19], the proposed benchmark aims at being aware of all the attendees along the whole video and detecting their eyeblinks at instance level. In summary, MPEblink is featured with multi-instance, uncon- strained, and untrimmed, which makes it more challenging and realistic than previous formulations. To perform eyeblink detection, previous methods [9, 10, 12, 19, 40] generally take a sequential approach contain- ing face detection, face tracking, and classification on the pre-extracted local eye clues within a temporal window. Whereas such a pipeline seems reasonable, it has several critical drawbacks. First, the use of isolated components may lead to sub-optimal results as they are not jointly op- timized and it is inefficient due to the inability to reuse the features of each stage. Second, the eyeblink features only contain the information from the pre-extracted local eye clues (i.e., a small part of face), lacking the useful infor- mation from global face context such as head pose and illu- mination condition that are crucial for detecting eyeblinks in the wild. Moreover, the pre-extracted local eye clues are also unreliable due to the localization challenge towards un- constrained scenarios. Third, the sequential approach leads to computational cost being sensitive to the subject amount, which is hard to meet the real-time running requirement. To tackle these issues, we propose a one-stage multi- person eyeblink detection framework called InstBlink, which can simultaneously detect human faces, track them over time, and do eyeblink detection at instance level. Inst- Blink takes inspiration from the existing query-based meth- ods [41, 45, 47, 48] and models the spatio-temporal face as well as eyeblink representations at instance level within each query. The insight is that the features can be effectively shared among these sub-tasks and the eyeblink features can be facilitated via face’s global contexts (e.g., head pose and illumination condition) with joint optimization and interac- tion, especially in unconstrained in-the-wild cases. Exper- iments verify the superiority of InstBlink in both effective- ness and efficiency, while also highlighting the critical chal- lenges of real-time multi-person eyeblink detection in the wild for untrimmed video. The main contributions of this work lie in 3 folders: •To our knowledge, it is the first time that instance-level multi-person eyeblink detection in untrimmed videos is for- mally defined and explored;•We introduce an unconstrained multi-person eyeblink detection dataset MPEblink that contains 686 untrimmed videos with 8748 eyeblink events, featured with more re- alistic and challenging. •We propose a one-stage multi-person eyeblink detec- tion method that can jointly perform face detection, track- ing, and instance-level eyeblink detection. Such a task-joint paradigm can benefit the sub-tasks uniformly.
Yu_Task_Residual_for_Tuning_Vision-Language_Models_CVPR_2023	Abstract Large-scale vision-language models (VLMs) pre-trained on billion-level data have learned general visual represen- tations and broad visual concepts. In principle, the well- learned knowledge structure of the VLMs should be inherited appropriately when being transferred to downstream tasks with limited data. However, most existing efficient transfer learning (ETL) approaches for VLMs either damage or are excessively biased towards the prior knowledge, e.g., prompt tuning (PT) discards the pre-trained text-based classifier and builds a new one while adapter-style tuning (AT) fully relies on the pre-trained features. To address this, we propose a new efficient tuning approach for VLMs named Task Resid- ual Tuning (TaskRes), which performs directly on the text- based classifier and explicitly decouples the prior knowledge of the pre-trained models and new knowledge regarding a target task. Specifically, TaskRes keeps the original classifier weights from the VLMs frozen and obtains a new classifier for the target task by tuning a set of prior-independent pa- rameters as a residual to the original one, which enables re- liable prior knowledge preservation and flexible task-specific knowledge exploration. The proposed TaskRes is simple yet effective, which significantly outperforms previous ETL methods (e.g., PT and AT) on 11 benchmark datasets while requiring minimal effort for the implementation. Our code is available at https://github.com/geekyutao/TaskRes.	1. Introduction Over the past decade, deep learning-based visual recogni- tion models [10, 18, 28, 55, 58] have achieved great success. These state-of-the-art models are often trained on a large amount of image and discrete label pairs. The discrete la- bel is generated by converting a detailed textual description, e.g., “American curl cat”, into a simple scalar, which strik- ingly eases the computation of loss function. However, this also results in two evident limitations: (i) the rich semantics *Equal contribution. †Corresponding author. 0.0 2.5 5.0 7.5 10.0 12.5 15.0 17.5 Number of Tunable Parameters (M)5860626466Score (%) Zero-shot CLIP CoOp CLIP-Adapter Tip-Adapter-F T askRes1-shot 2-shot 4-shot 8-shot 16-shotFigure 1. Performance comparison between Zero-shot CLIP [49], CoOp [73], CLIP-Adapter [15], Tip-Adapter-F [71] and our TaskRes on ImageNet with few shot settings. in the textual description are underused, and (ii) the trained models are limited to recognizing the close-set classes only. Recent large-scale vision-language model (VLM) pre- training [1, 23, 33, 49, 69] eliminates those limitations by learning visual representations via textual supervision. For instance, texts and images are encoded and mapped into a unified space via a contrastive loss during pre-training [49]. The pre-trained text encoder can then be used to synthesize a text-based classifier for image recognition given the corre- sponding natural language descriptions as shown in Figure 2 (a). Those pre-trained VLMs have demonstrated a powerful transferability on a variety of downstream tasks in a zero-shot manner. However, the effectiveness of the aforementioned models heavily relies on their large-scale architectures and training datasets. For instance, CLIP [49] has up to 428 mil- lion parameters and is trained on 0.4 billion text-image pairs, while Flamingo [1] boasts up to 80 billion parameters and is trained on a staggering 2.1 billion pairs. This makes it im- practical to fully fine-tune the model on downstream tasks in a low-data regime. For that reason, efficient transfer learning (ETL) [15, 71–73] on pre-trained VLMs has gained popularity. ETL represents transfer learning to downstream tasks in both This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 10899 𝐸!𝐸" “a phot of a [cat]”𝐶𝑜𝑠𝑖𝑛𝑒𝑆𝑖𝑚𝑖𝑙𝑎𝑟𝑖𝑡𝑦 cat𝐸!𝐸" 𝐶𝑜𝑠𝑖𝑛𝑒𝑆𝑖𝑚𝑖𝑙𝑎𝑟𝑖𝑡𝑦 cat𝒗𝟏, 𝒗𝟐, …, 𝒗[cat] 𝐸!𝐸" “a phot of a [cat]”𝐶𝑜𝑠𝑖𝑛𝑒𝑆𝑖𝑚𝑖𝑙𝑎𝑟𝑖𝑡𝑦 cat 𝐸!𝐸" “a phot of a [cat]”𝐶𝑜𝑠𝑖𝑛𝑒𝑆𝑖𝑚𝑖𝑙𝑎𝑟𝑖𝑡𝑦 cat𝑜𝑛𝑙𝑦𝑟𝑢𝑛𝑜𝑛𝑐𝑒𝑇𝑎𝑠𝑘𝑅𝑒𝑠𝑖𝑑𝑢𝑎𝑙(𝑖𝑛𝑑𝑒𝑝𝑒𝑛𝑑𝑒𝑛𝑡) (a) Zero-shot CLIP(b) Prompt Tuning (c) Adapter-style Tuning(d) Task Residual Tuning𝑒𝑥𝑡𝑟𝑎𝑎𝑟𝑐ℎ𝑖𝑡𝑒𝑐𝑡𝑢𝑟𝑒𝑟𝑢𝑛𝑒𝑣𝑒𝑟𝑦𝑡𝑖𝑚𝑒 𝑡𝑜𝑜𝑏𝑖𝑎𝑠𝑒𝑑 ✅ ❌ ✅ ❌ ❌… Figure 2. Illustration of (a) Zero-shot CLIP, (b) prompt tuning, (c) adapter-style tuning and (d) our proposed Task Residual Tuning (TaskRes). Our method introduces a prior-independent task residual to the fixed pre-trained classifier ( i.e., text embeddings of CLIP), being free of running the text encoder every time or extra architecture design. parameter- and data-efficient manner. The core of ETL is twofold: (i) properly inheriting the well-learned knowledge structure of VLMs, which is already transferable; (ii) effec- tively exploring the task-specific knowledge given limited data. However, most existing ETL approaches, i.e., prompt tuning (PT) [72, 73] and adapter-style tuning (AT) [15, 71], either damage the prior knowledge of VLMs or learn the new knowledge of a task in an inappropriate/insufficient way. For example, instead of using the pre-trained text- based classifier, CoOp [73] (in Figure 2 (b)) is proposed to learn a continuous prompt for synthesizing a completely new one, which inevitably causes the loss of previous knowl- edge. Consequently, CoOp underperforms Zero-shot CLIP by 1.03%/0.37% in 1-/2-shot learning on ImageNet (see Fig- ure 1). In contrast, CLIP-Adapter [15] preserves the pre- trained classifier, but is excessively biased towards the prior knowledge when learning a new task, i.e., it transforms the pre-trained classifier weights to be task-specific as illustrated in Figure 2 (c). This results in an inferior new knowledge exploration, thereby a lower accuracy as shown in Figure 1. For better ETL on pre-trained VLMs, we propose a new efficient tuning approach named Task Residual Tuning (TaskRes), which performs directly on the text-based classi- fier and explicitly decouples the old knowledge of the pre- trained models and the new knowledge for a target task. The rationale is that the decoupling enables a better old knowledge inheritance from VLMs and a more flexible task- specific knowledge exploration ,i.e., the learned knowledge w.r.t. the task is independent on the old knowledge. Specif- ically, TaskRes keeps the original classifier weights frozen and introduces a set of prior-independent parameters that are added to the weights. These additive parameters, tuned for adaptation to the target task, are thus named “task residual”. To gain insight into how TaskRes works, we perform ex-tensive experiments across 11 benchmark datasets [73] and conduct a systematic investigation of learned task residu- als. The experimental results demonstrate that introducing task residual can significantly enhance the transfer perfor- mance. We visualize the correlation between the magnitude of learned task residual and the difficulty of transferring a pre-trained model to a downstream task, and observe that the magnitude increases with the transfer difficulty. This sug- gests the residual is automatically adapted to the task to fully explore the new knowledge, thereby achieving a new state- of-the-art performance on 11 diverse datasets. Furthermore, it is worth noting that our method requires minimal effort for the implementation, i.e., technically adding one line of code only. Our contributions are summarized below: • We for the first time emphasize the necessity of a proper knowledge inheritance from pre-trained VLMs to downstream tasks via ETL, reveal the pitfalls of exist- ing tuning paradigms, and conduct an in-depth analysis to manifest that decoupling the old pre-trained knowl- edge and the new task-specific knowledge is the key. • We propose a new efficient tuning approach named Task Residual Tuning (TaskRes), which achieves a better old knowledge inheritance from VLMs and a more flexible task-specific knowledge exploration. • TaskRes is convenient for use, which needs a few tuning parameters and effortless implementation.
Zhang_Prototypical_Residual_Networks_for_Anomaly_Detection_and_Localization_CVPR_2023	Abstract Anomaly detection and localization are widely used in industrial manufacturing for its efficiency and effectiveness. Anomalies are rare and hard to collect and supervised mod- els easily over-fit to these seen anomalies with a handful of abnormal samples, producing unsatisfactory performance. On the other hand, anomalies are typically subtle, hard to discern, and of various appearance, making it difficult to detect anomalies and let alone locate anomalous regions. To address these issues, we propose a framework called Prototypical Residual Network (PRN), which learns feature residuals of varying scales and sizes between anomalous and normal patterns to accurately reconstruct the segmen- tation maps of anomalous regions. PRN mainly consists of two parts: multi-scale prototypes that explicitly represent the residual features of anomalies to normal patterns; a multi- size self-attention mechanism that enables variable-sized anomalous feature learning. Besides, we present a variety of anomaly generation strategies that consider both seen and unseen appearance variance to enlarge and diversify anoma- lies. Extensive experiments on the challenging and widely used MVTec AD benchmark show that PRN outperforms cur- rent state-of-the-art unsupervised and supervised methods. We further report SOTA results on three additional datasets to demonstrate the effectiveness and generalizability of PRN.	1. Introduction The human cognition and visual system has an inherent ability to perceive anomalies [53]. Not only can humans distinguish between defective and non-defective images, but they can also point to the location of anomalies even if they have seen none or only a limited number of anomalies. Anomaly detection (image-level binary classification) and anomaly localization (pixel-level binary classification) are introduced for the same purpose, and have been widely used ∗Corresponding author. InputGTPatchCoreDRAOursFigure 1. Anomaly detection and localization examples on MVTec [4]. Compared with the unsupervised method PatchCore [41] and the supervised method DRA [13], the proposed PRN is able to locate the anomalous regions more accurately. in various scenarios due to their efficiency and remarkable accuracy, including industrial defect detection [4, 7, 34, 61], medical image analysis [52] and video surveillance [32]. Given its importance, a significant amount of work has been devoted to anomaly detection and anomaly localization, but few have addressed both detection and localization prob- lems well at the same time. We argue that real-world anoma- lous data weaken these models mainly in three aspects: I) the amount of abnormal samples is limited and significant fewer than normal samples, producing data distributions that lead to a naturally imbalanced learning problem; II) anomalies are typically subtle and hard to discern, since normal patterns still dominate the anomalous image; identifying abnormal regions out of the whole image is the key to anomaly detec- tion and localization; III) the appearance of anomalies varies significantly, i.e., abnormal regions can take on a variety of sizes, shapes and numbers, and such appearance variations This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 16281 InputGTDRADevNet Ours Figure 2. Indecipherable problem of supervised methods DevNet [35] and DRA [13]. Both images are detected as anomalous. Other methods mistakenly highlight normal regions rather than defect regions, whereas PRN correctly pinpoints the defect regions. make it challenging to well-localizing all the anomalies. Without adequate anomalies for training, unsupervised models become the de facto dominant approaches, which get rid of the imbalance problem by learning the distribu- tion of normal samples [5, 9, 10, 12, 18, 25, 41–43, 47, 67] or generating sufficient synthetic anomalies [26, 28, 50, 63, 68]. However, these methods are opaque to genuine anomalies, resulting in implicit decisions that may induce many false negatives and false positives. Besides, unsupervised methods rely heavily on the quality of normal samples, and thus are not robust enough and perform poorly on uncalibrated or noisy datasets [19]. As shown in Fig. 1, unsupervised models predict broad regions around the anomaly. We attribute this problem to less discriminative abilities of these methods. Recently, several supervised methods [13, 35, 46] are in- troduced. DeepSAD [46] enlarges the margin between the anomaly and the one-class center in the latent space to obtain more compact one-class descriptors by limit seen anomalies. DRA [13] and DevNet [35] formulate anomaly detection as a multi-instance learning (MIL) problem, scoring an image as anomaly if any image patch is a defect region. MIL-based methods enforce the learning at fine-grained image patch level, which effectively reduces the interference of normal patches in the anomalous images. Yet, these approaches typi- cally struggle to accurately locate all anomalous regions with image-level supervision, as shown in Fig. 1. In particular, when the anomalous regions only occupy a tiny part of image patches, image-level representation may be dominated by the normal regions and disregards tiny anomalous, which may cause inconsistent image-level and pixel level perfor- mance as shown in Table 1. Furthermore, as shown in Fig. 2, these methods also encounter uninterpretable problems when making decisions. In this paper, we propose a framework called Prototypical Residual Network (PRN) as an effective remedy for afore- said issues on anomaly detection and localization. First, we propose multi-scale prototypes to represent normal patterns. In contrast to previous methods for constructing normal pat-terns from concatenated feature memory [41] or random sampled feature maps [63], we construct normal patterns with prototypes of intermediate feature maps of different scales, thereby preserving the spatial information and provid- ing precise and representative normal patterns. Further, we obtain the feature map residuals via the deviation between the anomalous image and the closest prototype at each scale, and we add multi-scale fusion blocks to exchange informa- tion across different scales. Second, since the appearance of anomaly regions varies a lot, it is necessary to learn relation- ships among patches from multiple receptive fields. Thus, we introduce a multi-size self-attention [33, 55, 58, 59] mecha- nism, which operates on patches of different receptive fields to detect patch-level inconsistencies at different sizes. Fi- nally, unlike previous methods [13, 35] that use image-level supervision for training, our model learns to reconstruct the anomaly segmentation map with pixel-level supervision, which focuses more on the anomalous regions and preserves better generalization. Besides, we put forward a variety of anomaly generation strategies that efficiently mitigate the impact of data imbalance and enrich the anomaly appearance. With the proposed modules, our method achieves more accu- rate localization than previous unsupervised and supervised methods, as shown in Fig. 1 and Fig. 2. The main contributions of this paper are summarized as follows: •We propose a novel Prototypical Residual Networks for anomaly detection and localization. Equipped with multi-scale prototypes and the multi-size self-attention mechanism, PRN learns residual representations among multi-scale feature maps and within the multi-size re- ceptive fields at each scale. •We present a variety of anomaly generation strategies that considering both seen and unseen appearance vari- ance to enlarge and diversify anomalies. •We perform extensive experiments on four datasets to show that our approach achieves new SOTA anomaly detection performance and outperforms current SOTA in anomaly localization performance by a large margin.
Yun_IFSeg_Image-Free_Semantic_Segmentation_via_Vision-Language_Model_CVPR_2023	Abstract Vision-language (VL) pre-training has recently gained much attention for its transferability and flexibility in novel concepts (e.g., cross-modality transfer) across various visual tasks. However, VL-driven segmentation has been under- explored, and the existing approaches still have the bur- den of acquiring additional training images or even seg- mentation annotations to adapt a VL model to downstream segmentation tasks. In this paper, we introduce a novel image-free segmentation task where the goal is to perform semantic segmentation given only a set of the target seman- tic categories, but without any task-specific images and an- notations. To tackle this challenging task, our proposed method, coined IFSeg, generates VL-driven artificial image- segmentation pairs and updates a pre-trained VL model to a segmentation task. We construct this artificial train- ing data by creating a 2D map of random semantic cat- egories and another map of their corresponding word to- kens. Given that a pre-trained VL model projects visual and text tokens into a common space where tokens that share the semantics are located closely, this artificially generated word map can replace the real image inputs for such a VL model. Through an extensive set of experiments, our model not only establishes an effective baseline for this novel task but also demonstrates strong performances compared to ex- isting methods that rely on stronger supervision, such as task-specific images and segmentation masks. Code is avail- able at https://github.com/alinlab/ifseg .	1. Introduction Understanding a new concept with less cost ( e.g., col- lecting data, annotations, or training) is a challenging yet essential problem in machine learning [41]. The most com- mon practice is fine-tuning a foundation model, pre-trained on a large amount of data [3,6,12,18], for downstream tasks. *Equal contribution †Work was done while at KAIST “grass” word“cat” word“dog” word“other” wordFigure 1. Visualization of image-free segmentation results via IFSeg on a web image. Here, we present a web image ( Top) and its segmentation results ( Middle andBottom ) of our image-free segmentation approach. Note that our model is not trained with any task-specific images and annotations, but only the text words ( e.g., “grass”, “cat”, “dog” and “other”) as semantic categories. In particular, such large-scale models have shown success- ful adaptation to downstream tasks with only little supervi- sion across vision [6] and language [3] domains. Recently, pre-training approaches in the vision-language (VL) domain have also achieved remarkable results in transferring to novel tasks ( e.g., few-shot or zero-shot transfer [37]) with various elaborate designs, including modality interaction between the dual encoders [20, 32], the multi-modal encoder [22, 43], and the encoder-decoder [1, 8, 39, 42, 44, 49]. Semantic segmentation is one of the crucial tasks in com- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 2967 puter vision that requires understanding dense representa- tions for pixel-wise classifications. Inspired by the success of the contrastive VL pre-training, CLIP [32], several re- cent attempts [15, 25, 27, 48, 53] have explored CLIP-based segmentation approaches for better transferability ( e.g., zero- shot [4, 45] and open-vocabulary segmentation [51]). How- ever, the existing zero-shot or open-vocabulary segmentation approaches still suffer from a burden of training on addi- tional image data, segmentation annotations [15, 25, 48, 53], or natural language supervision [27,47], to adapt pre-trained VL models to downstream segmentation tasks. In the wild, however, such training data is not readily available; e.g., there would be no task-specific training images or labels for novel web images like Fig. 1. This limitation inspires us to investigate how to fully utilize the VL models for seman- tic segmentation in a lightweight manner, even without any image data or human-annotated supervision. Meanwhile, the recent encoder-decoder VL models [1, 8, 39, 42, 44, 49] also have gained popularity with their unique characteristics of image-to-text generation via the VL de- coder network. Motivated by this, we explore the potential usability of the VL decoder to segment pixels in the text generation manner as an alternative to traditional vision segmentation decoders,
Zhang_Differentiable_Architecture_Search_With_Random_Features_CVPR_2023	Abstract Differentiable architecture search (DARTS) has signif- icantly promoted the development of NAS techniques be- cause of its high search efficiency and effectiveness but suf- fers from performance collapse. In this paper, we make efforts to alleviate the performance collapse problem for DARTS from two aspects. First, we investigate the expres- sive power of the supernet in DARTS and then derive a new setup of DARTS paradigm with only training Batch- Norm. Second, we theoretically find that random features dilute the auxiliary connection role of skip-connection in supernet optimization and enable search algorithm focus on fairer operation selection, thereby solving the performance collapse problem. We instantiate DARTS and PC-DARTS with random features to build an improved version for each named RF-DARTS and RF-PCDARTS respectively. Experi- mental results show that RF-DARTS obtains 94.36% test ac- curacy on CIFAR-10 (which is the nearest optimal result in NAS-Bench-201), and achieves the newest state-of-the-art top-1 test error of 24.0% on ImageNet when transferring from CIFAR-10. Moreover, RF-DARTS performs robustly across three datasets (CIFAR-10, CIFAR-100, and SVHN) and four search spaces (S1-S4). Besides, RF-PCDARTS achieves even better results on ImageNet, that is, 23.9% top-1 and 7.1% top-5 test error, surpassing representative methods like single-path, training-free, and partial-channel paradigms directly searched on ImageNet.	1. Introduction Differentiable architecture search (DARTS) [27] has demonstrated both higher search efficiency and better search efficacy than early pioneering neural architecture search (NAS) [1,50,51] attempts in the image classification task. In the past few years, many following works further *Equal contributions.This work is done during Yonggang Li’s in- ternship at MEGVII Technology. This work is supported by Science and Technology Innovation 2030-New Generation Artificial Intelligence (2020AAA0104401).improve DARTS by introducing additional modules, such as Gumbel-softmax [13], early stop criterion [23], auxiliary skip-connection [10], etc. We have witnessed tremendous improvements in the image recognition task, but it is getting farther away from exploring how DARTS works. Newly, in this work, we intend to demystify DARTS by disassembling key modules rather than make it more complex. We overview the vanilla DARTS paradigm, and sum- mary three key modules, namely dataset ,evaluation met- ric, and supernet as follows: •Dataset. DARTS [27] searches on proxy dataset and then transfers to target dataset due to huge requirements for GPU memory. PC-DARTS [36] proves that proxy datasets inhibit the effectiveness of DARTS and directly searching on target dataset ob- tains more promising architectures. UnNAS [25] and RLNAS [46] ablate the role of labels in DARTS, and further conclude that ground truth labels are not neces- sary for DARTS. •Evaluation metric. DARTS [27] introduces architec- ture parameters to reflect the strengths of the candidate operations. PT-DARTS [32] suspects the effectiveness of architecture parameters and shows that the magni- tude of architecture parameters does not necessarily indicate how much the operation contributes to the su- pernet’s performance. FreeNAS [45] and TE-NAS [7] further put forward training-free evaluation metrics to predict the performance of candidate architectures. •Supernet. DARTS encodes all candidate architectures in search space into the supernet. The search cell of supernet will change as the search space changes. R- DARTS [41] proposes four challenging search spaces S1-S4 where DARTS obtains inferior performance than the Random-search baseline. R-DARTS attributes the failure of vanilla DARTS to the dominant skip- connections. Thus R-DARTS concludes that the topol- ogy of supernet has great influence on the efficacy of DARTS. P-DARTS [9] finds that the depth gap This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 16060 (a) Results on CIFAR-10 (b) Results on CIFAR-100 (c) Results on ImageNet16-120 Figure 1. The correlation between the DARTS supernet performance (blue histograms) and the searched architecture performance (orange histograms) in NAS-Bench-201 [14] search space (best viewed in color). We directly searched on target datasets CIFAR-10, CIFAR-100, and ImageNet16-120 in three runs and plot average results. Histograms on three datasets reflect a consistent phenomenon. Supernet performance and the expressive power are positively correlated. However, supernet with higher performance can not search for better architectures, and supernet optimized with only BN has the best search effect. between search and evaluation architecture prevents DARTS from achieving better search results. Option ( state )Learnable modules in supernet Expressive Convolution BN Affine Power A (unexplored ) ! ! strong B (explored ) ! %wwww C (unexplored ) % ! D (unexplored ) % % weak Table 1. Enumerate four combinations of learnable modules in supernet. We keep the composition of supernet unchanged and just change the optimizable weights. !means updating weights with the optimizer. %means freezing weights at the initialization. In this paper, we take a further step to investigate the su- pernet in DARTS from two respects. First, we ablate the expressive power of supernet in DARTS. Specifically, each convolution operation (like separable convolutions) consists of two learnable modules: convolution layers and Batch- Norm (BN) [21] layers. Each convolution layer is followed by a private BN layer. To study the expressive power in iso- lation, we thoroughly traverse four combinations (named A, B, C, and D for simple) of learnable modules as shown in Tab. 1. Existing DARTS variants [9, 27, 36] adopt option B by default, which disables BN affine weights and only trains convolution weights during the supernet training. On the contrary, option A, C, and D are still unexplored, thus it is a mystery what will happen when supernet is equipped with different expressive power, that is, trained with option of A, C, and D. Hence, we make a sanity check between su- pernet performance and searched architecture performance across the above four combinations. As shown in Fig. 1, the relative ranking of supernet performance is A ≈B>C>D, which is consistent with the ranking of supernet’s expres- sive power. However, the relative ranking of searched archi-tecture performance is C ≫D>A≈B, which sounds count- intuitive. This result implies that the performance of su- pernet is not such significant and scaling random fea- tures with BN affine weights is good enough for archi- tecture search . Therefore, we propose a new extension of DARTS with random features driven by the surprising re- sults of only training BN. Second, we explore the work- ing mechanism of random features by considering skip- connection roles in DARTS. Skip-connection in DARTS plays two roles [10]: 1) as a shortcut to help the optimization of supernet, and 2) a candidate operation for architecture search. Random features dilute the role of auxiliary con- nection that skip-connection plays in supernet training and enable DARTS to focus on fairer operation selection, thus implicitly solving performance collapse of DARTS [10,41]. Based on the versatility of supernet optimization, we arm popular DARTS [27] and PC-DARTS [36] with ran- dom features to build more effective algorithms RF-DARTS and RF-PCDARTS. On CIFAR-10, RF-DARTS obtains 94.36% test accuracy that is the nearest optimal re- sults (94.37%) in NAS-Bench-201. RF-DARTS achieves the state-of-the-art 24.0% top-1 test error on ImageNet when transferred from CIFAR-10 in DARTS search space. RF-DARTS also performs robustly on CIFAR-10, CIFAR- 100, and SVHN across S1-S4. RF-PCDARTS directly searches on ImageNet and achieves 23.9% top-1 test er- ror, which surpasses representative methods from single- path, training-free, and partial channel paradigms. Overall, comprehensive results reveal that the expressive power of DARTS supernet is over-powerful, and random features is just perfect for DARTS. We hope these essential analyses and results will inspire new understandings for NAS. 16061
Xu_MM-3DScene_3D_Scene_Understanding_by_Customizing_Masked_Modeling_With_Informative-Preserved_CVPR_2023	Abstract Masked Modeling (MM) has demonstrated widespread success in various vision challenges, by reconstructing masked visual patches. Yet, applying MM for large-scale 3D scenes remains an open problem due to the data spar- sity and scene complexity. The conventional random mask- ing paradigm used in 2D images often causes a high risk of ambiguity when recovering the masked region of 3D scenes. To this end, we propose a novel informative-preserved re- construction, which explores local statistics to discover and preserve the representative structured points, effectively en- hancing the pretext masking task for 3D scene understand- ing. Integrated with a progressive reconstruction man- ner, our method can concentrate on modeling regional ge- ometry and enjoy less ambiguity for masked reconstruc- tion. Besides, such scenes with progressive masking ra- *Equal contribution. †Corresponding authors.tios can also serve to self-distill their intrinsic spatial con- sistency, requiring to learn the consistent representations from unmasked areas. By elegantly combining informative- preserved reconstruction on masked areas and consistency self-distillation from unmasked areas, a unified framework called MM-3DScene is yielded. We conduct comprehensive experiments on a host of downstream tasks. The consistent improvement ( e.g., +6.1% [email protected] on object detection and +2.2% mIoU on semantic segmentation) demonstrates the superiority of our approach.	1. Introduction 3D scene understanding plays an essential role in vari- ous visual applications, such as virtual reality, robot naviga- tion, and autonomous driving. Over the past few years, deep learning has dominated 3D scene parsing tasks [26, 40, 73]. However, traditional supervised learning methods require massive annotation of 3D scene data that are extremely la- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 4380 borious to obtain [9], where millions of points or mesh ver- tices per scene need to be labeled. To solve this, self-supervised learning (SSL) becomes a favorable choice since it can extract rich representa- tions without any annotation [10, 16, 17]. Masked Mod- eling (MM) [16, 57], as one of the representative meth- ods in SSL, recently draws significant attention in the vi- sion community. Recently, It has been explored in 3D vi- sion [31,39,51,67,69,70], where these 3D MM approaches randomly mask local regions of point clouds, and pre-train neural networks to reconstruct the masked areas. Neverthe- less, such random masking paradigms are not feasible for large-scale 3D scenes, which often causes a high risk of re- construction ambiguity. As illustrated in Fig. 1 (a), a chair and a TV are totally masked, which are extremely difficult to be recovered without any context guidance. Such ambi- guity often makes MM difficult to learn informative repre- sentation for 3D scenes. Hence, we ask a natural question: can we customize a better way of masked modeling for 3D scene understanding? To tackle this question, we propose a novel informative- preserved masked reconstruction scheme in this paper. Specifically, we leverage local statistics of each point ( i.e., the difference between each point and its neighboring points in terms of color and shape) as guidance to discover the rep- resentative structured points which are usually located at the boundary regions in the 3D scene. We denote these points as ‘Informative Points ’ since they provide highly useful information hints and rich semantic context for assisting masked reconstruction. To this end, our mask strategy is definite: to preserve Informative Points in a scene and mask other points. In this way, the basic geometric information of a scene is explicitly retained , which effectively simpli- fies the pretext task and reduces ambiguity. Based on our mask strategy, a progressive masked recon- struction manner is integrated, to better model the masked areas. As illustrated in Fig. 1 (b), during each iteration, our method concentrates on reconstructing the local regional geometric patterns rather than rebuilding the original intact scene. In doing so, it enjoys less ambiguity and is able to restore accurate geometric information. Moreover, we realize the information of unmasked ar- eas ( i.e., Informative Points) is underexplored. We find that there exists point correspondence in the unmasked ar- eas under progressive masking ratios. Accordingly, we in- troduce a dual-branch encoding scheme for learning such intrinsic consistency, with the ultimate goal of unearthing the consistent ( i.e., masking-invariant) representations from unmasked areas. This leads to a more powerful SSL frame- work on 3D scenes, called MM-3DScene, which elegantly combines the masked modeling on the masked and un- masked areas in 3D scenes together, while complements each other. It achieves superior performance in Table 7 (v).Datasets Complexity Task Gain (from scratch) S3DIS Entire floor, office segmentation (+1.5%) mIoU segmentation (+2.2%) mIoU ScanNet Large rooms detection (+4.4) [email protected] detection (+6.1) [email protected] SUN-RGBD Cluttered rooms detection (+2.9) [email protected] detection (+4.4) [email protected] Table 1. Summary of fine-tuning MM-3DScene on various downstream tasks and datasets for 3D understanding. Our MM- 3DScene conspicuously boosts the performance of the baseline model trained from scratch. Our contributions are motivated and comprehensive: • We raise the concept of Informative Points – the points providing significant information hints, and indicate that preserving them is critical for assisting masked modeling on 3D scenes (Table 8). • For masked areas, we propose an informative- preserved reconstruction scheme to focus on restoring the regional geometry in a novel progressive manner, which explicitly simplifies the pretext task. • For unmasked areas, we introduce a self-distillation branch, which is encouraged to learn spatial-consistent representations under progressive masking ratios. • A unified self-supervised framework, called MM- 3DScene, delivers performance improvements across on various downstream tasks and datasets (Table 1).
Zhang_GrowSP_Unsupervised_Semantic_Segmentation_of_3D_Point_Clouds_CVPR_2023	Abstract We study the problem of 3D semantic segmentation from raw point clouds. Unlike existing methods which primarily rely on a large amount of human annotations for training neural networks, we propose the first purely unsupervised method, called GrowSP , to successfully identify complex se- mantic classes for every point in 3D scenes, without need- ing any type of human labels or pretrained models. The key to our approach is to discover 3D semantic elements via progressive growing of superpoints. Our method consists of three major components, 1) the feature extractor to learn per-point features from input point clouds, 2) the superpoint constructor to progressively grow the sizes of superpoints, and 3) the semantic primitive clustering module to group superpoints into semantic elements for the final semantic segmentation. We extensively evaluate our method on mul- tiple datasets, demonstrating superior performance over all unsupervised baselines and approaching the classic fully- supervised PointNet. We hope our work could inspire more advanced methods for unsupervised 3D semantic learning.	1. Introduction Giving machines the ability to automatically discover se- mantic compositions of complex 3D scenes is crucial for many cutting-edge applications. In the past few years, there has been tremendous progress in fully-supervised semantic segmentation for 3D point clouds [14]. From the seminar works PointNet [40] and SparseConv [12] to a plethora of *Corresponding authorrecent neural models [21, 27, 41, 55, 60], both the accuracy and efficiency of per-point semantic estimation have been greatly improved. Unarguably, the success of these meth- ods primarily relies on large-scale human annotations for training deep neural networks. However, manually annotat- ing real-world 3D point clouds is extremely costly due to the unstructured data format [3, 20]. To alleviate this prob- lem, a number of recent methods start to use fewer 3D point labels [19, 69], cheaper 2D image labels [59, 77], or active annotations [22,63] in training. Although achieving promis- ing results, they still need tedious human efforts to annotate or align 3D points across images for particular datasets, thus being inapplicable to novel scenes without training labels. In this paper, we make the first step towards unsuper- vised 3D semantic segmentation of real-world point clouds. To tackle this problem, there could be two strategies: 1) to na¨ıvely adapt existing unsupervised 2D semantic segmen- tation techniques [4, 7, 24] to 3D domain, and 2) to apply existing self-supervised 3D pretraining techniques [17, 66] to learn discriminative per-point features followed by clas- sic clustering methods to obtain semantic categories. For unsupervised 2D semantic methods, although achieving en- couraging results on color images, they can be hardly ex- tended to 3D point clouds primarily because: a) there is no general pretrained backbone to extract high-quality fea- tures for point clouds due to the lack of representative 3D datasets akin to ImageNet [46] or COCO [29], b) they are usually designed to group pixels with similar low-level fea- tures, e.g. colors or edges, as a semantic class, whereas such a heuristic is normally not satisfied in 3D point clouds due to point sparsity and spatial occlusions. For self-supervised 3D pretraining methods, although the pretrained per-point This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 17619 features could be discriminative, they are lack of seman- tic meanings fundamentally because the commonly adopted data augmentation techniques do not explicitly capture cat- egorical information. Section 4 clearly demonstrates that all these methods fail catastrophically on 3D point clouds. Given a sparse point cloud composed of multiple seman- tic categories, we can easily observe that a relative small local point set barely contains distinctive semantic infor- mation. Nevertheless, when the size of a local point set is gradually growing, that surface patch naturally emerges as a basic element or primitive for a particular semantic class, and then it becomes much easier for us to identify the cat- egories just by combining those basic primitives. For ex- ample, two individual 3D points sampled from a spacious room are virtually meaningless, whereas two patches might be easily identified as the back and/or armof chairs. Inspired by this, we introduce a simple yet effective pipeline to automatically discover per-point semantics, sim- ply by progressively growing the size of per-point neigh- borhood, without needing any human labels or pretrained backbone. In particular, our architecture consists of three major components: 1) a per-point feature extractor which is flexible to adopt an existing (untrained) neural network such as the powerful SparseConv [12]; 2) a superpoint con- structor which progressively creates larger and larger su- perpoints during training to guide semantic learning; 3) a semantic primitive clustering module which aims to group basic elements of semantic classes via an existing clustering algorithm such as K-means. The key to our pipeline is the superpoint constructor together with a progressive growing strategy in training. Basically, this component drives the feature extractor to progressively learn similar features for 3D points within a particular yet grow ingsuperpoint, while the features of different superpoints tend to be pushed as distinct elements of semantic classes. Our method is called GrowSP and Figure 1 shows qualitative results of an indoor 3D scene. Our contributions are: • We introduce the first purely unsupervised 3D semantic segmentation pipeline for real-world point clouds, with- out needing any pretrained models or human labels. • We propose a simple strategy to progressively grow su- perpoints during network training, allowing meaningful semantic elements to be learned gradually. • We demonstrate promising semantic segmentation results on multiple large-scale datasets, being clearly better than baselines adapted from unsupervised 2D methods and self-supervised 3D pretraining methods. Our code is at: https://github.com/vLAR-group/GrowSP
Yu_PEAL_Prior-Embedded_Explicit_Attention_Learning_for_Low-Overlap_Point_Cloud_Registration_CVPR_2023	Abstract Learning distinctive point-wise features is critical for low-overlap point cloud registration. Recently, it has achieved huge success in incorporating Transformer into point cloud feature representation, which usually adopts a self-attention module to learn intra-point-cloud features first, then utilizes a cross-attention module to perform fea- ture exchange between input point clouds. The advan- tage of Transformer models mainly benefits from the use of self-attention to capture the global correlations in fea- ture space. However, these global correlations may in- volve ambiguity for point cloud registration task, espe- cially in indoor low-overlap scenarios, because the corre- lations with an extensive range of non-overlapping points may degrade the feature distinctiveness. To address this is- sue, we present PEAL, a Prior-embedded Explicit Attention Learning model. By incorporating prior knowledge into the learning process, the points are divided into two parts. One includes points lying in the putative overlapping region and the other includes points located in the putative non- overlapping region. Then PEAL explicitly learns one-way attention with the putative overlapping points. This simplis- tic design attains surprising performance, significantly re- lieving the aforementioned feature ambiguity. Our method improves the Registration Recall by 6+% on the challenging 3DLoMatch benchmark and achieves state-of-the-art per- formance on Feature Matching Recall, Inlier Ratio, and Registration Recall on both 3DMatch and 3DLoMatch.	1. Introduction Rigid point cloud registration has always been a foun- dational yet challenging task in 3D vision and robotics [2, 3, 10, 25], which aims to estimate an optimal rigid trans- formation to align two point clouds. Benefiting from the superior feature representation of deep networks, keypoints-based point cloud registration methods have become dominant in recent years [4, 9, 12, ∗Corresponding author: Wenhui Zhou, [email protected]; Yu Zhang, [email protected] Overlap: 15.6% # Patch Corr: 256 Inlier ratio: 10.9% Overlap: 15.6% # Patch Corr: 256 Inlier ratio: 74.6% Overlap: 15.6% # Point Corr : 500 Inlier ratio: 85.4% Overlap: 15.6% # Point Corr : 500 Inlier ratio: 16.4% (a) Implicit attention, Patch Feature and Correspondences (c) Explicit attention, Patch Feature and Correspondences(b) Implicit attention, Point Correspondences (d) Explicit attention, Point CorrespondencesFigure 1. Given two low-overlap point clouds, PEAL adopts an explicit attention learning fashion and learns discriminative su- perpoint (patch) features (c), which results in significant higher patch and point inlier ratios. In contrast, GeoTransformer learns ambiguous patch features (a). For example, PEAL is able to ac- curately identify corresponding chairs among multiple chairs and distinguish them from the floor and table, while GeoTransformer mismatches them. Zoom in for details. 34, 37]. The core idea is to learn to match the learned key- points across different point clouds. Recently, the keypoint- free methods [25, 36] demonstrate promising performance following the coarse-to-fine fashion. They seek correspon- dences between downsampled point clouds (superpoints), which are then propagated to individual points to yield dense correspondences. Thus, the accuracy of superpoint matching is crucial to the overall performance of point cloud registration. GeoTransformer [25] proposes a geo- metric self-attention module that encodes the distance of point pairs and the angle of triplet to extract transformation- invariant features. This approach significantly improves the accuracy of superpoint matching. However, GeoTransformer may still suffer from am- biguous matching in certain scenarios with numerous sim- ilar structure or low geometric discriminative patches (su- perpoints) [25]. Moreover, the self-attention mechanism may exacerbate matching ambiguity, especially for low- overlap registration tasks. Prior works [3, 25] advocate that modeling geometric consistent correlations among overlap- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 17702 ping superpoints/points is the key to the success of super- points/points matching, while the global correlations learn- ing via the geometric self-attention is inevitably interfered by numerous superpoints in the non-overlapping region. In other words, the correlations with non-overlapping super- points may disrupt the inter-frame geometric consistency learning and degrade the feature distinctiveness for registra- tion, which makes the resultant learned superpoint features ambiguous and leads to numerous outlier matches (Fig. 1 (a)). To address the aforementioned issues, we design a Prior- embedded Explicit Attention Learning model (PEAL). It first leverages an overlap prior to divide the superpoints into anchor ones (the superpoints lying in putative overlapping region) and non-anchor ones (the superpoints located in pu- tative non-overlapping region). Then it alleviates the inter- ference of non-anchor superpoints by introducing an one- way attention mechanism, which solely models the correla- tions from non-anchor superpoints to anchor ones. Benefit- ting from the promising overlap ratio in the anchor region, anchor superpoints can be reckoned as simultaneously ex- isting in both two frames, thus the one-way attention is ca- pable of acquiring the essential local geometric consistent correlation from the anchor region, which helps the non- anchor superpoints encoding the inter-frame local geomet- ric consistency and relieves the global feature ambiguity (Fig. 1 (c) ). Furthermore, the embedding prior design in- volved in PEAL makes refining transformation possible in an iterative fashion. In this paper, we introduce two models depending on the methods of obtaining prior, with extensive experiments on indoor benchmarks demonstrating the superiority of PEAL. Compared to state-of-the-art methods, both of the two mod- els achieve significant improvements on Registration Recall on the challenging 3DLoMatch benchmark. In summary, our contributions are summarized as follows: • To the best of our knowledge, we are the first to explic- itly inject overlap prior into Transformer to facilitate low-overlap point cloud registration, and various over- lap priors can be integrated into this framework, such as 3D overlap prior, 2D overlap prior, and self-overlap- prior. • An explicit one-way attention module, which can sig- nificantly relieves the feature ambiguity generated by self-attention. It can be plugged into other transformer- based point cloud registration networks. • A novel iterative pose refined fashion for low-overlap point cloud registration.
Yang_TopDiG_Class-Agnostic_Topological_Directional_Graph_Extraction_From_Remote_Sensing_Images_CVPR_2023	Abstract Rapid development in automatic vector extraction from remote sensing images has been witnessed in recent years. However, the vast majority of existing works concentrate on a speciﬁc target, fragile to category variety, and hardly achieve stable performance crossing different categories. In this work, we propose an innovative class-agnostic model, namely TopDiG, to directly extract topological directional graphs from remote sensing images and solve these issues. Firstly, TopDiG employs a topology-concentrated node de- tector (TCND) to detect nodes and obtain compact percep- tion of topological components. Secondly, we propose a dynamic graph supervision (DGS) strategy to dynamical- ly generate adjacency graph labels from unordered nodes. Finally, the directional graph (DiG) generator module is designed to construct topological directional graphs from predicted nodes. Experiments on the Inria ,CrowdAI , GID ,GF2 andMassachusetts datasets empirically demonstrate that TopDiG is class-agnostic and achieves competitive performance on all datasets.	1. Introduction Vector maps that are represented as topological direc- tional graphs act as the foundation to various remote sens- ing applications, such as property mapping, cartographic generalization and disaster assessment [21, 25]. Traditional manual or semi-automatic vector map generation from re- mote sensing images is extremely time-consuming and ex- pensive. In contrast, state-of-the-art approaches, including segmentation-based [4, 9, 13, 30], contour-based [1, 16, 22, 29, 35, 39] and graph generation [3, 26, 31–34, 41] methods have typically developed to achieve automation. Howev- er, these works are concentrated on a speciﬁc category and can hardly achieve satisfactory performance when applied to other classes. Among aforementioned approaches, a dominant paradigm is the segmentation-based method. It follows the (a) PolyWorld (b) Enhanced iCurb (c) Ours Figure 1. Visual illustrations of current and our approaches on different targets . In contrast with PolyWorld and Enhanced iCurb which only serve a speciﬁc category, TopDiG is class-agnostic and can tackle both polygon-shape and line-shape targets. Yellow dots refer to detected nodes while blue arrowed lines indicate the direc- tional topological connection between node pairs (best viewed by zooming in). segmentation-vectorization pipeline and requires sophis- ticated vectorization procedures to binary masks. Typical examples that fall in this scope include PolyMapper [14], Frame Field [9] and ASIP [13]. These methods retrieve coarse raster maps with missing details and unavoidably de- mand elaborate post-processing. Another paradigm mainly adopts contour-based instance segmentation approaches, which usually reﬁne initial contours to obtain vector maps. For instance, Polygon-RNN [6], Polygon-RNN++ [1], Deep Snake [22], SharpContour [39] and E2EC [35] can delineate the outlines of polygon-shape targets, such as buildings and water bodies. However, it depends on the quality of initial contours and can barely be reliable on line-shape targets, such as roads. The ﬂourishing graph generation methods construct the topological graph based on nodes and their connectivity. A few of such approaches, This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 1265 including RoadTracer [3], VecRoad [26], iCurb [33] and RNGDet [32] focus on line-shape targets by iteratively predicting nodes in a one-by-one manner. Nevertheless, these methods suffer from the low efﬁciency, the accumu- lated node connectivity error, and the poor reliability to polygon-shape targets. Alternatively, the connectivity of the nodes also can be recovered from adjacency matrix as introduced in PolyWorld [41] and csBoundary [31]. These workﬂows successfully produce visually pleasing vector topology without irregular edges and overly smoothed corners. Unfortunately, for intricate or sinuous structures, such methods lead to severe topological errors. Given the class-dependent characteristics, existing work- s can hardly apply to other classes as illustrated in Fig- ure 1. For example, PolyWolrd [41] (Figure 1(a)) is able to extract well-vectorized buildings but fails to delineate road networks. By contrast, Enhanced iCurb [34], orig- inally concerning line-shape road curbs, is challenged by polygon-shape buildings (Figure 1(b)). They adopt the sim- ilar scheme where topological graphs are constructed by connecting detected nodes and aim at either polygon-shape or line-shape targets, respectively. Nevertheless, neither of them can achieve stable and reliable topological directional graphs regardless of varying categories. In this work, we propose a class-agnostic approach named TopDiG which can robustly obtain precise topological directional graphs both for polygon-shape and line-shape targets (Figure 1(c)). The underlying innovation is that the TopDiG formulates diverse topological structures as directional graphs and nar- rows the gap among categorical varieties. Besides, we fur- ther develop a dynamic graph supervision strategy that en- ables ﬂexible arrangement of the predicted nodes and stabi- lizes performance crossing different categories. Our contri- butions are summarized as follows: ADynamic Graph Supervision (DGS) strategy is de- signed to generate the ground truth of adjacency matrix in an on-the-ﬂy manner during training. Instead of utiliz- ing the adjacency matrix labels established from ordered ground truth nodes [31, 36, 41], we dynamically generate such labels according to real-time unordered predict nodes in each training epoch. Our strategy alleviates the compul- sory assumption that the sequence of predict nodes must be in consistent with real ones as in PolyWorld [41]. Conse- quently, DGS can facilitate the connectivity of unordered nodes and ease the demand for the accurate positions of nodes. We further propose a novel topology-concentrated node detector (TCND) to guarantee an appropriate density of predicted nodes. Unlike PolyWorld [41] and csBound- ary [31] that mainly employ semantic contexts, TCND con- centrates on compact geometric textures via the meticu- lous perception of topological components, which boosts the topological APLS score by approximately 8.06 %. AClass-agnostic Topological Directional Graph Extrac-tion (TopDiG) approach is proposed to extract polygon- shape and line-shape targets, i.e., buildings, water bodies and roads, from remote sensing images. In contrast with existing approaches that can only serve a speciﬁc catego- ry, TopDiG directly performs class-independent vector map generation from diverse targets. We introduce TCND and directional graph (DiG) generator module to retain the ge- ometrical shapes, i.e., polygon-shape and line-shape tar- gets. Our method is performed in an end-to-end manner and does not require initial contours or additional post- processing. The TopDiG outperforms the segmentation- based, contour-based and previous graph generation ap- proaches, achieving a competitive performance with bound- ary mIoU scores of 68.39 %, 72.51 %, 74.51 %and 75.28 % onInria ,CrowdAI ,GID andGF2 datasets, respective- ly. Moreover, TopDiG can construct reliable topological di- rectional graphs, with the application to Massachusetts dataset, achieving an average path length score ( APLS ) of 64.60 %.
Yang_Directional_Connectivity-Based_Segmentation_of_Medical_Images_CVPR_2023	Abstract Anatomical consistency in biomarker segmentation is crucial for many medical image analysis tasks. A promising paradigm for achieving anatomically consistent segmentation via deep networks is incorporating pixel connectivity, a basic concept in digital topology, to model inter-pixel relationships. However, previous works on connectivity modeling have ignored the rich channel-wise directional information in the latent space. In this work, we demonstrate that effective disentanglement of directional sub-space from the shared latent space can significantly enhance the feature representation in the connectivity-based network. To this end, we propose a directional connectivity modeling scheme for segmentation that decouples, tracks, and utilizes the directional information across the network. Experiments on various public medical image segmentation benchmarks show the effectiveness of our model as compared to the state-of-the-art methods. Code is available at https://github.com/Zyun-Y/DconnNet.	1. Introduction Maintaining anatomical consistency in the segmentation of medical images is important but challenging, as minor geometric errors may change the global topology [1, 2] and cause functional mistakes in downstream clinical decision-making [3]. Anatomical consistency in images can be expressed with topological properties, such as pixel connectivity and adjacency [4, 5]. As such, by directly modeling the mutual information between pixels or regions, graph-based methods have long been used to correct topological and geometrical errors [6-8]. However, such classic machine vision techniques usually depend on manually defined priors and thus are not easily generalizable for a wide variety of applications. Alternative to the classic approaches, deep learning-based segmentation methods utilized an encoder-decoder architecture [9] to learn from a group of pixels in a particular receptive field at each layer. More recently, significant progress has been made in capturing the inter- pixel dependency inside a network’s latent space [10-12]; Figure 1. The latent space differences between traditional pixel-classification-based and connectivity-based models. In the former, only categorical features, e.g., boundaries, are highlighted; while in the latter, the feature map also contains directional information (e.g., the horizontal connections between boundary pixels). Figure 2. The flows of the two groups of latent features (categorical and directional) in the latent space of DconnNet, are visualized by T-SNE [13]. They were first disentangled (Sec 3.2) and then effectively fused in a projected shared manifold (Sec 3.3). The colors are rendered based on the results of clustering. however, very few studies have been conducted on the problem modeling side of the networks. A typical segmentation network models the problem as a pure pixel-wise classification task and uses a segmentation mask as the only label. Yet, this pixel-wise modeling scheme is suboptimal as it does not directly exploit inter-pixel relationships and geometrical properties [14, 15]. Thus, these models may result in low spatial coherence (i.e., inconsistent predictions for neighboring pixels that share similar spatial features) in their prediction [16]. Especially, when applied to high noise/artifacts medical data, the lower spatial consistency may lead to topological issues [17]. The concept of pixel connectivity has long been used to ensure the basic topological duality of separation and connectedness in digital images [18]. More recently, in the context of deep learning, the connectivity masks, reviewed Directional Connectivity-based Segmentation of Medical Images Ziyun Yang Sina Farsiu Duke University Duke University Durham, NC, United States Durham, NC, United States [email protected] [email protected] This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 11525 Figure 3. Illustration of generating the connectivity mask from the segmentation mask by traversing 8-neighbor pixel connectivity. For each pixel, channel 𝑪𝒊’s binary value in the converted connectivity vector carries the categorical information (connected or not connected) while 𝒊 encodes the directional information (direction of connection). in Section 2.1, have been introduced as a topological extension of the segmentation mask [15]. Using connectivity masks as training labels has several advantages over segmentation masks. In terms of problem modeling, using a connectivity mask inherently changes the problem from pixel-wise classification to connectivity prediction, which models and enhances the topological representation between pixels of interest. In terms of label representation, a connectivity mask is more informative in three ways: first, a connectivity mask stores the categorical information among the connections of pixels and it is inter-pixel relation-aware; second, it sparsely represents edge pixels [16]; third, it contains rich directional information channel-wisely. Thus, a network trained with connectivity masks has both categorical (reflected by connectivity) and directional features in its latent space, each of which forms a specific sub-latent space, as shown in Fig. 1. In previous studies [15, 16, 19-22], these two groups of features were learned simultaneously through a shared network path which may result in highly coupled latent space and introduce redundancy [23]. Further, effectively disentangling meaningful subspaces from the shared latent space has been shown effective in accounting for the dependencies/independencies between features [24, 25]. Inspired by the idea of latent space disentanglement, in this paper, we propose a novel directional connectivity-based segmentation network (DconnNet) to disentangle the directional subspace from the shared latent space and utilize the extracted directional features to enhance the overall data representation, as in Fig. 2. The disentangling process is conducted by a sub-path slicing-based module called Sub-path Direction Excitation (SDE). The directional-based feature enhancement is applied in a coarse-to-fine manner using an Interactive Feature-space Decoder (IFD) with two top-down interactive decoding flows. Finally, we propose a novel Size Density loss (SDL) that alleviates the common data imbalance problem in medical datasets with a label size distribution-based weighting scheme. With experiments on different public medical image analysis benchmarks, we demonstrate the superiority of DconnNet against other state-of-art methods.
Zhang_PointCert_Point_Cloud_Classification_With_Deterministic_Certified_Robustness_Guarantees_CVPR_2023	Abstract Point cloud classiﬁcation is an essential component in many security-critical applications such as autonomous driving and augmented reality. However, point cloud classi- ﬁers are vulnerable to adversarially perturbed point clouds. Existing certiﬁed defenses against adversarial point clouds suffer from a key limitation: their certiﬁed robustness guar- antees are probabilistic, i.e., they produce an incorrect cer- tiﬁed robustness guarantee with some probability. In this work, we propose a general framework, namely PointCert, that can transform an arbitrary point cloud classiﬁer to be certiﬁably robust against adversarial point clouds with de- terministic guarantees. PointCert certiﬁably predicts the same label for a point cloud when the number of arbitrarily added, deleted, and/or modiﬁed points is less than a thresh- old. Moreover, we propose multiple methods to optimize the certiﬁed robustness guarantees of PointCert in three appli- cation scenarios. We systematically evaluate PointCert on ModelNet and ScanObjectNN benchmark datasets. Our re- sults show that PointCert substantially outperforms state- of-the-art certiﬁed defenses even though their robustness guarantees are probabilistic.	1. Introduction Point cloud classiﬁcation [ 11,29,30,38,44,50] has many safety-critical applications, including but not limited to, au- tonomous driving and augmented reality. However, vari- ous studies [ 12,16,22,32,40,43,46,51,52] showed that point cloud classiﬁcation is vulnerable toadversarial point clouds. In particular, an attacker can carefully add, delete, and/or modify a small number of points in a point cloud to make it misclassiﬁed by a point cloud classiﬁer. Existing defenses against adversarial point clouds can be categorized intoempirical defenses[ 7,21,33,41,46,49,53] andcertiﬁed defenses[ 5,6,23]. The key limitation of em- pirical defenses is that they cannot provide formal guaran-tees, and thus are often broken by advanced, adaptive at- tacks [ 34]. Therefore, we focus on certiﬁed defenses in this work. Randomized smoothing [ 5] and PointGuard [ 23] are two state-of-the-art certiﬁed defenses against adversar- ial point clouds. In particular, randomized smoothing adds random noise (e.g., Gaussian noise) to a point cloud, while PointGuard randomly subsamples a point cloud. Due to the randomness, their certiﬁed robustness guarantees areprob- abilistic, i.e., they produce incorrect robustness guarantees with some probability (callederror probability). For in- stance, when the error probability is 0.001, they produce in- correct robustness guarantees for 1 out of 1,000 point-cloud classiﬁcations on average. Such probabilistic guarantees are insufﬁcient for security-critical applications that frequently classify point clouds. In this work, we propose PointCert, theﬁrst certiﬁed de- fense that hasdeterministicrobustness guarantees against adversarial point clouds. PointCert can transform an arbi- trary point cloud classiﬁerf(calledbase point cloud classi- ﬁer) to be certiﬁably robust against adversarial point clouds. Speciﬁcally, given a point cloud and a base point cloud clas- siﬁerf, PointCertﬁrst divides the point cloud into multiple disjoint sub-point clouds using a hash function, then usesf to predict a label for each sub-point cloud, andﬁnally takes a majority vote among the predicted labels as the predicted label for the original point cloud. We prove that PointCert certiﬁably predicts the same label for a point cloud when the number of arbitrarily added, deleted, and/or modiﬁed points is no larger than a threshold, which is known ascertiﬁed perturbation size. Moreover, we also prove that our derived certiﬁed perturbation size istight, i.e., without making as- sumptions on the base point cloud classiﬁerf, it is theoret- ically impossible to derive a certiﬁed perturbation size for PointCert that is larger than ours. We consider three scenarios about how PointCert could be applied in practice and propose methods to optimize the performance of PointCert in these scenarios. In particular, we consider two parties:model providerandcustomer. A model provider (e.g., Google, Meta) has enough labeled This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 9496 data and computation resource to train a base point cloud classiﬁerfand shares it with customers (e.g., a less re- sourceful company). Givenf, a customer uses PointCert to classify its (adversarial) point clouds. We note that the model provider and customer can be the same entity, e.g., a company trains and usesfitself. We consider three scenar- ios, in whichfis trained by the model provider differently and/or used by a customer differently. Scenario I represents a naive application of PointCert, in which the base point cloud classiﬁerfis trained using a standard training algorithm and a customer directly applies PointCert to classify its point clouds based onf. PointCert achieves suboptimal performance in Scenario I becausef, trained on point clouds, is not accurate at classifying sub- point clouds as they have different distributions. Therefore, in Scenario II, we consider a model provider trainsfto optimize the performance of PointCert. In particular, the model provider divides each training point cloud into multi- ple sub-point clouds following PointCert and trainsfbased on sub-point clouds. In Scenario III, we consider the model provider has trainedfusing a standard training algorithm (like Scenario I). However, instead of directly applyingfto classify sub-point clouds, a customer prepends aPoint Com- pletion Network (PCN)[ 48] tof. Speciﬁcally, a PCN takes a sub-point cloud as input and outputs a completed point cloud, which is then classiﬁed byf. Moreover, we propose a new loss function to train the PCN such that its completed point clouds are classiﬁed byfwith higher accuracy, which further improves the performance of PointCert. We perform systematic evaluation on ModelNet40 dataset [ 1] and two variants of ScanObjectNN dataset [ 2]. Our experimental results show that PointCert signiﬁcantly outperforms the state-of-the-art certiﬁed defenses (random- ized smoothing [ 5] and PointGuard [ 23]) even though their robustness guarantees are probabilistic. For instance, on ModelNet40 dataset, PointCert achieves acertiﬁed accu- racyof 79% when an attacker can arbitrarily perturb at most 50 points in a point cloud, where certiﬁed accuracy is a lower bound of testing accuracy. Under the same setting, the certiﬁed accuracy of both randomized smoothing and PointGuard is0. We also extensively evaluate PointCert in the three application scenarios. In summary, we make the following contributions: (1) We propose PointCert, theﬁrst certiﬁed defense with de- terministic robustness guarantees against adversarial point clouds. (2) We design multiple methods to optimize the performance of PointCert in multiple application scenarios. (3) We extensively evaluate PointCert and compare it with state-of-the-art certiﬁed defenses.
Yu_Learning_Procedure-Aware_Video_Representation_From_Instructional_Videos_and_Their_Narrations_CVPR_2023	Abstract The abundance of instructional videos and their narra- tions over the Internet offers an exciting avenue for un- derstanding procedural activities. In this work, we pro- pose to learn video representation that encodes both ac- tion steps and their temporal ordering, based on a large- scale dataset of web instructional videos and their narra- tions, without using human annotations. Our method jointly learns a video representation to encode individual step con- cepts, and a deep probabilistic model to capture both tem- poral dependencies and immense individual variations in the step ordering. We empirically demonstrate that learn- ing temporal ordering not only enables new capabilities for procedure reasoning, but also reinforces the recognition of individual steps. Our model significantly advances the state-of-the-art results on step classification (+2.8%/+3.3% on COIN / EPIC-Kitchens) and step forecasting (+7.4% on COIN). Moreover, our model attains promising re- sults in zero-shot inference for step classification and fore- casting, as well as in predicting diverse and plausible steps for incomplete procedures. Our code is available at https://github.com/facebookresearch/ProcedureVRL.	1. Introduction Many of our daily activities ( e.g. cooking or crafting) are highly structured, comprising a set of action steps con- ducted in a certain ordering. Yet how these activities are performed varies among individuals. Consider the exam- ple of making scrambled eggs as shown in Fig. 1. While most people tend to whisk eggs in a bowl, melt butter in a pan, and cook eggs under medium heat, expert chefs have recommended to crack eggs into the pan, add butter, and stir them under high heat. Imagine a vision model that can account for the individual variations and reason about the temporal ordering of action steps in a video, so as to infer prior missing steps, recognize the current step, and forecast *Work done while Yiwu Zhong was an intern at Meta. †Co-corresponding authors. Video 1Video 2whisk eggs in bowlcook eggscrack egg in panadd butterStep Descriptions flatten doughbake pizzabake cookiesStep ForecastingStep Classification knead dough Our ModelOur ModelFigure 1. Top: During training, our model learns from procedu- ral videos and step descriptions to understand individual steps and capture temporal ordering and variations among steps. Bottom : Once trained, our model supports zero-shot step classification and forecasting, yielding multiple credible predictions. a future step. Such a model will be immensely useful for a wide range of applications including augmented reality, virtual personal assistant, and human-robot interaction. Understanding complex procedural activities has been a long-standing challenge in the vision community [7, 18, 22, 39, 41, 46]. While many prior approaches learn from anno- tated videos following a fully supervised setting [13,27,69], this paradigm is difficult to scale to a plethora of activ- ities and their variants among individuals. A promising solution is offered by the exciting advances in vision-and- language pre-training, where models learn from visual data (images or videos) and their paired text data (captions or narrations) [30, 43, 54, 70] in order to recognize a variety of concepts. This idea has recently been explored to analyze instructional videos [33, 37], yet existing methods are lim- ited to recognize single action steps in procedural activities. In this paper, we present a first step towards modeling temporal ordering of action steps in procedural activities by learning from instructional videos and their narrations. Our key innovation lies in the joint learning of a video repre- sentation aiming to encode individual step concepts, and a deep probabilistic model designed to capture temporal de- pendencies and variations among steps. The video represen- tation, instantiated as a Transformer network, is learned by This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 14825 matching a video clip to its corresponding narration. The probabilistic model, built on a diffusion process, is tasked to predict the distribution of the video representation for a missing step, given steps in its vicinity. With the help of a pre-trained vision-and-language model [43], our model is trained using only videos and their narrations from auto- matic speech recognition (ASR), and thus does not require any manual annotations. Once learned, our model celebrates two unique bene- fits thanks to our model design and training framework. First, our model supports zero-shot inference given an input video, including the recognition of single steps and fore- casting of future steps, and can be further fine-tuned on downstream tasks. Second, our model allows sampling mul- tiple video representations when predicting a missing action step, with each presenting a possibly different hypothesis of the step ordering. Instead of predicting a single represen- tation with the highest probability, sampling from a proba- bilistic model provides access to additional high-probability solutions that might be beneficial to prediction tasks with high ambiguity or requiring user interactions. We train our models on a large-scale instructional video dataset collected from YouTube (HowTo100M [38]), and evaluate them on two public benchmarks (COIN [55] and EPIC-Kitchens-100 [10]) covering a wide range of pro- cedural videos and across the tasks of step classification and step forecasting. Through extensive experiments, we demonstrate that (1) our temporal model is highly effec- tive in forecasting future steps, outperforming state-of-the- art methods by a large margin of +7.4% in top-1 accu- racy on COIN; (2) modeling temporal ordering reinforces video representation learning, leading to improved clas- sification results ( +2.8% /+3.3% for step classification on COIN/EPIC-Kitchens) when probing the learned represen- tations; (3) our training framework offers strong results for zero-shot step classification and forecasting; and (4) sam- pling from our probabilistic model yields diverse and plau- sible predictions of future steps. Contributions . Our work presents the first model that leverages video-and-language pre-training to capture the temporal ordering of action steps in procedural activities. Our key technical innovation lies in the design of a deep probabilistic model using a diffusion process, in tandem with video-and-language representation learning. The re- sult is a model and a training framework that establish new state-of-the-art results on both step classification and fore- casting tasks across the major benchmarks. Besides, our model is capable of generating diverse step predictions and supports zero-shot inference.
Yin_GIVL_Improving_Geographical_Inclusivity_of_Vision-Language_Models_With_Pre-Training_Methods_CVPR_2023	Abstract A key goal for the advancement of AI is to develop technologies that serve the needs not just of one group but of all communities regardless of their geographical re- gion. In fact, a signiﬁcant proportion of knowledge is lo- cally shared by people from certain regions but may not apply equally in other regions because of cultural dif- ferences. If a model is unaware of regional character- istics, it may lead to performance disparity across re- gions and result in bias against underrepresented groups. We propose GIVL, a Geographically Inclusive Vision-and- Language Pre-trained model. There are two attributes of geo-diverse visual concepts which can help to learn geo- diverse knowledge: 1) concepts under similar categories have unique knowledge and visual characteristics, 2) con- cepts with similar visual features may fall in completely different categories. Motivated by the attributes, we de- sign new pre-training objectives Image-Knowledge Match- ing (IKM) and Image Edit Checking (IEC) to pre-train GIVL. Compared with similar-size models pre-trained with similar scale of data, GIVL achieves state-of-the-art (SOTA) and more balanced performance on geo-diverse V&L tasks.Code and data are released at https://github.com/ WadeYin9712/GIVL .	1. Introduction Vision-Language Pre-trained Models (VLPs) [ 9,23,24, 29,53] have achieved remarkable performance on Vision- Language (V&L) tasks including visual question answer- ing [ 11,12,15], image-text retrieval [ 22], and image cap- tioning [ 19,27]. Pre-trained with large-scale corpora of image-text pairs, e.g. COCO [ 27], OpenImages [ 21]. VLPs are capable of learning multi-modal representations and can be effectively ﬁne-tuned on downstream V&L tasks. While VLPs can solve a broad range of V&L tasks, to deploy VLPs in real-world applications, it is essential to consider the geographical inclusivity1of VLPs. Because of geographic differences, images from different regions em- body a large amount of knowledge that is locally shared but cannot be applied in other regions, i.e. geographically di- verse. For example, in Figure 1, the festivals in different 1We use regions as a proxy to estimate inclusivity of V&L models. People in the same regions may have different cultures and traditions. This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 10951 regions look different. Ideally, a geographically inclusive VLP should be capa- ble of achieving comparable performance over all the im- ages, regardless of their origins. However, current VLPs does not perform equally well on data from different re- gions. For example, prior works [ 28,49] show that on geo- diverse V&L tasks, there is nearly a 20% performance dis- crepancy between Western and East Asian images when current VLPs are applied. To combat such geographical bias, we aim to design methods to make VLPs achieve more balanced performance across regions. One solution to mitigating bias is to obtain diverse task- speciﬁc annotations for each region and ﬁne-tune VLPs on the new annotations. However, according to [ 17], most Amazon MTurk annotators are from US and India, and may be unfamiliar with the cultures of other regions. Thus, it is unrealistic to obtain large-scale geo-diverse annotations even in such a popular crowdsourcing platform. Pre-training a uniﬁed VLP with large-scale unannotated geo-diverse images and corresponding knowledge could make the VLP a foundation to provide more generalizable representations and help to transfer on comprehending im- ages from various regions easier. In this paper, we propose GIVL ,aGeographically Inclusive Vision-and- Language Pre-trained model. We focus on how to encourage GIVL to better learn geo-diverse knowledge on images from different regions during its pre-training stage . We observe two attributes of geo-diverse visual concepts that can contribute to learning geo-diverse knowledge: A1: Concepts under similar categories have unique knowledge and visual characteristics. For example, tra- ditional Western and Chinese festivals, like Christmas and Chinese New Year in Figure 1, are held with different rit- uals and their decoration style differs as well. It is neces- sary for GIVL to learn the difference between their corre- sponding knowledge and precisely distinguish these visual concepts. On the other hand, Christmas andChinese New Year are both festivals. Learning the commonalities of vi- sual concepts (e.g., both images in Figure 1belong to the same category “festival”) would help model connect West- ern and non-Western concepts and contribute to more effec- tive transfer on geo-diverse images. A2: Concepts with similar visual features may lie in completely different categories. In Figure 2,Chinese pa- per cuttings share visual features (e.g., color, shape) with red frisbee . Similarly. sugar cane andﬂute share visual fea- tures. However, these concepts are not related to each other. Since geo-diverse images cover a broader range of visual concepts, differentiating visually similar concepts given vi- sual contexts is also essential. To this end, besides common objectives Masked Lan- guage Modeling (MLM) and Image-Text Matching (ITM) for pre-training VLPs, we propose two additional pre- (a)(b) Figure 2. Example of Chinese paper cuttings and red frisbee (left), sugar cane and ﬂute (right). Different concepts may be visually similar, but they may have completely different functionalities. training objectives, Image-Knowledge Matching (IKM) andImage Edit Checking (IEC) . IKM is used to learn the alignment between images and corresponding textual knowledge in Wikipedia. It requires GIVL to not only judge if the input textual knowledge matches input images, but also identify whether the visual concepts described in input knowledge falls into similar categories of the concepts in in- put images. This encourages GIVL to learn corresponding relationship between knowledge and images as well as rec- ognize similarity among geo-diverse visual concepts. IEC is proposed to identify whether a visual concept in input image is replaced by another concept that is visually similar but lies in an irrelevant category (see Fig. 3for an example). It enables GIVL to capture nuances between visually simi- lar concepts after the replacement given visual contexts. Our contributions and empirical results are as follows: •By considering the attributes of geo-diverse visual con- cepts, we propose two novel V&L pre-training objec- tives Image-Knowledge Matching (IKM) and Image Edit Checking (IEC) that can greatly improve the geo- graphical inclusivity of VLPs. •Compared with similar-size VLPs pre-trained with similar scale of data, GIVL achieves state-of-the-art (SOTA) and more balanced performance over dif- ferent regions on geo-diverse V&L tasks including MaRVL [ 28], GD-VCR [ 49] and WIT Image-Text Re- trieval [ 38]. For geo-diverse zero-shot image classi- ﬁcation on Dollar Street dataset2, GIVL outperforms VinVL [ 53] 26%.
Zhang_Seeing_a_Rose_in_Five_Thousand_Ways_CVPR_2023	Abstract What is a rose, visually? A rose comprises its intrinsics, including the distribution of geometry, texture, and material specific to its object category. With knowledge of these in- trinsic properties, we may render roses of different sizes and shapes, in different poses, and under different lighting condi- tions. In this work, we build a generative model that learns to capture such object intrinsics from a single image, such as a photo of a bouquet. Such an image includes multiple instances of an object type. These instances all share the same intrinsics, but appear different due to a combination of variance within these intrinsics and differences in extrinsic factors, such as pose and illumination. Experiments show that our model successfully learns object intrinsics (distri- bution of geometry, texture, and material) for a wide range of objects, each from a single Internet image. Our method achieves superior results on multiple downstream tasks, in- cluding intrinsic image decomposition, shape and image generation, view synthesis, and relighting. “Nature!... Each of her works has an essence of its own; each of her phenomena a special characterisation; and yet their diversity is in unity. ” – Georg Christoph Tobler	1. Introduction The bouquet in Figure 1 contains many roses. Although each rose has different pixel values, we recognize them as individual instances of the same type of object. Such un- derstanding is based on the fact that these instances share the same object intrinsics —the distributions of geometry, texture, and material that characterize a rose. The difference in appearance arises from both the variance within these distributions and extrinsic factors such as object pose and environment lighting. Understanding these aspects allows us to imagine and draw additional, new instances of roses with varying shape, pose, and illumination. In this work, our goal is to build a model that captures such object intrinsics from a single image , and to use this model for shape and image generation under novel view- points and illumination conditions, as illustrated in Figure 1. This problem is challenging for three reasons. First, we only have a single image. This makes our work fundamen- tally different from existing works on 3D-aware image gen- eration models [8, 9, 27, 28], which typically require a large dataset of thousands of instances for training. In comparison, 1“People where you live ... grow five thousand roses in one garden ... And yet what they’re looking for could be found in a single rose.” — Quote from The Little Prince by Antoine de Saint-Exup ´ery. This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 962 the single image contains at most a few dozen instances, making the inference problem highly under-constrained. Second, these already limited instances may vary signif- icantly in pixel values. This is because they have different poses and illumination conditions, but neither of these factors are annotated or known. We also cannot resort to existing tools for pose estimation based on structure from motion, such as COLMAP [35], because the appearance variations violate the assumptions of epipolar geometry. Finally, the object intrinsics we aim to infer are proba- bilistic, not deterministic: no two roses in the natural world are identical, and we want to capture a distribution of their geometry, texture, and material to exploit the underlying multi-view information. This is therefore in stark contrast to existing multi-view reconstruction or neural rendering meth- ods for a fixed object or scene [25, 26, 40]. These challenges all come down to having a large hypothesis space for this highly under-constrained problem, with very limited visual observations or signals. Our solution to address these chal- lenges is to design a model with its inductive biases guided byobject intrinsics . Such guidance is two-fold: first, the instances we aim to present share the same object intrinsics, or the same distribution of geometry, texture, and material; second, these intrinsic properties are not isolated, but inter- weaved in a specific way as defined by a rendering engine and, fundamentally, by the physical world. Specifically, our model takes the single input image and learns a neural representation of the distribution over 3D shape, surface albedo, and shininess of the object, factoring out pose and lighting variations, based on a set of instance masks and a given pose distribution of the instances. This ex- plicit, physically-grounded disentanglement helps us explain the instances in a compact manner, and enables the model to learn object intrinsics without overfitting the limited obser- vations from only a single image. The resulting model enables a range of applications. For example, random sampling from the learned object intrinsics generates novel instances with different identities from the input. By modifying extrinsic factors, the synthesized in- stances can be rendered from novel viewpoints or relit with different lighting configurations. Our contributions are three-fold: 1.We propose the problem of recovering object intrinsics, including both 3D geometry, texture, and material prop- erties, from just a single image of a few instances with instance masks. 2.We design a generative framework that effectively learns such object intrinsics. 3.Through extensive evaluations, we show that the model achieves superior results in shape reconstruction and generation, novel view synthesis, and relighting.Object VarianceUnknown PosesRe- lighting3D- Aware SinGAN [36] ✓ ✓ NeRF [26] ✓ D-NeRF [32] ✓ ✓ GNeRF [25] ✓ ✓ Neural-PIL [7] ✓ ✓ NeRD [6] ✓ ✓ EG3D [9] ✓ ✓ Ours ✓ ✓ ✓ ✓ Table 1. Comparisons with prior works. Unlike existing 3D-aware generative models, our method learns from a very limited number of observations. Unlike multi-view reconstruction methods, our method models variance among observations from training inputs.
Xu_Low-Light_Image_Enhancement_via_Structure_Modeling_and_Guidance_CVPR_2023	Abstract This paper proposes a new framework for low-light im- age enhancement by simultaneously conducting the appear- ance as well as structure modeling. It employs the struc- tural feature to guide the appearance enhancement, lead- ing to sharp and realistic results. The structure model- ing in our framework is implemented as the edge detection in low-light images. It is achieved with a modified gen- erative model via designing a structure-aware feature ex- tractor and generator. The detected edge maps can accu- rately emphasize the essential structural information, and the edge prediction is robust towards the noises in dark ar- eas. Moreover, to improve the appearance modeling, which is implemented with a simple U-Net, a novel structure- guided enhancement module is proposed with structure- guided feature synthesis layers. The appearance model- ing, edge detector, and enhancement module can be trained end-to-end. The experiments are conducted on represen- tative datasets (sRGB and RAW domains), showing that our model consistently achieves SOTA performance on all datasets with the same architecture. The code is available at https://github.com/xiaogang00/SMG-LLIE.	1. Introduction The low-light enhancement aims to recover normal-light and noise-free images from dark and noisy pictures, which is a long-standing and significant computer vision topic. It has broad application fields, including low-light imag- ing [11, 27, 50], and also benefits many downstream vision tasks, e.g., nighttime detection [28, 48, 64]. Some meth- ods have been proposed to tackle the low-light enhancement problem. They design networks that learn to manipulate color, tone, and contrast [7, 10, 45, 63], and some recent works also account for noise in images [24, 55]. Most of these works optimize the appearance distance between the output and the ground truth. However, they ignore the ex- plicit modeling of structural details in dark areas and thus *Corresponding author. 15.00 18.00 21.00 24.00 0.600 0.800 1.000 PSNR SSIMLOL-synthetic (sRGB) 16.00 19.00 22.00 25.00 0.600 0.700 0.800 0.900 PSNR SSIMLOL-real (sRGB) 16.00 19.00 22.00 0.520 0.570 0.620 0.670 PSNR SSIMSID (sRGB) 26.00 28.00 30.00 0.720 0.770 0.820 PSNR SSIMSID (RAW) Legend for sRGB Legend for RAW Figure 1. Our method consistently achieves SOTA performance on different sRGB/RAW datasets with the same network architecture. resulting in blurry outcomes and low SSIM [51] values, as shown in Fig. 2. Some works [34, 74] have noticed the ef- fect of employing structural information, e.g., edge, to pro- mote the enhancement. Edge guides the enhancement by distinguishing between different parts in the dark regions. Moreover, adding sensible edge priors into dark regions re- duces the ill-posed degree in optimizing the appearance re- construction. These frameworks [34, 74] perform the struc- ture modeling with encoder-decoder-based networks and a regression loss. However, the corresponding structure mod- eling results are not satisfying due to the uncertainty in the dark areas caused by severely poor visibility and noise. Fur- thermore, the strategy of using the extracted structural infor- mation needs to be improved from the existing straightfor- ward concatenation approach [34, 74]. In this paper, we propose to utilize a generative model S trained with a GAN loss to perform the structure modeling with the form of edges. Then, we design a new mechanism Eto facilitate the initial low-light appearance enhancement (the module is denoted as A) with structure-guided feature synthesis. With effective structure modeling and guidance, our framework can output sharp and realistic results with satisfactory reconstruction quality as shown in Fig. 2. Compared with previous structure modeling networks, This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 9893 (a) Input (b) Structure of (a) (c) SNR (CVPR 2022) (d) Structure Modeling (e) Ours (f) Ground Truth Figure 2. A challenging low-light frame (a), from SID-sRGB [3], enhanced by a SOTA method (c) and our method (e). Our method can synthesize the structure map (d) from the input image, leading to clearer details, more distinct contrast, and more vivid color. Al- though (c) has high PSNR as 28.17, its SSIM is low as 0.75. Ours achieves high scores for both dB and SSIM, as 28.60dB and 0.80. the proposed generative model Shas two significant modi- fications. First, we notice the impact of providing structure- aware descriptors into both the encoder and decoder of S, disentangling the appearance representation and underlin- ing the structural information. Thus, we design a Structure- Aware Feature Extractor (SAFE) as the encoder part, which extracts structure-aware features from the dark image and its gradients via spatially-varying operations (achieved with adaptive long-range and short-range computations). The extracted structure-aware tensors are then fed into the de- coder part to generate the desired structure maps. Moreover, different from current approaches, which employ the struc- ture maps of normal-light images to conduct the regression learning, we find the nice property of using a GAN loss. The GAN loss can reduce the artifacts in the generated structure maps that are caused by the noise and invisibility, highlight- ing the essential structure required for enhancement. The backbone of Sis implemented as a modified StyleGAN. To boost the appearance by leveraging the obtained structure maps, we design a Structure-Guided Enhancement Module (SGEM) as E. The main target of SGEM is to learn the residual, which can improve the initial appearance mod- eling results. In SGEM, spatially-adaptive kernels and nor- malization parameters are generated according to the struc- ture maps. Then, the features in each layer of the SGEM’s decoder will be processed with spatially-adaptive convolu- tions and normalization. Although the overall architecture of SGEM takes the form of a simple U-Net [38], it can ef- fectively enhance the original appearance. S,A, and Ecan be trained end-to-end simultane- ously. Extensive experiments are conducted on representa- tive benchmarks. Experimental results show that our frame- work achieves SOTA performance on both PSNR and SSIMmetrics with the same architecture on all datasets, as shown in Fig. 1. In summary, our work’s contribution is four-fold. 1. We propose a new framework for low-light enhance- ment by conducting structure modeling and guidance simultaneously to boost the appearance enhancement. 2. We design a novel structure modeling method, where structure-aware features are formulated and trained with a GAN loss. 3. A novel structure-guided enhancement approach is proposed for appearance improvement guided by the restored structure maps. 4. Extensive experiments are conducted on different datasets in both sRGB and RAW domains, showing the effectiveness and generalization of our framework.
Yang_Improving_Visual_Grounding_by_Encouraging_Consistent_Gradient-Based_Explanations_CVPR_2023	Abstract We propose a margin-based loss for tuning joint vision- language models so that their gradient-based explanations are consistent with region-level annotations provided by humans for relatively smaller grounding datasets. We re- fer to this objective as Attention Mask Consistency (AMC) and demonstrate that it produces superior visual ground- ing results than previous methods that rely on using vision- language models to score the outputs of object detectors. Particularly, a model trained with AMC on top of standard vision-language modeling objectives obtains a state-of-the- art accuracy of 86.49% in the Flickr30k visual grounding benchmark, an absolute improvement of 5.38% when com- pared to the best previous model trained under the same level of supervision. Our approach also performs exceed- ingly well on established benchmarks for referring expres- sion comprehension where it obtains 80.34% accuracy in the easy test of RefCOCO+, and 64.55% in the difficult split. AMC is effective, easy to implement, and is general as it can be adopted by any vision-language model, and can use any type of region annotations.	1. Introduction Vision-language pretraining using images paired with captions has led to models that can transfer well to an ar- ray of tasks such as visual question answering, image-text retrieval and visual commonsense reasoning [6,18,22]. Re- markably, some of these models are also able to perform vi- sual grounding by relying on gradient-based explanations. While Vision-Language Models (VLMs) take advantage of the vast amounts of images and text that can be found on the web, carefully curated data with grounding annotations in the form of boxes, regions, or segments is consider- ably more limited. Our work aims to improve the ground- ing or localization capabilities of vision-language models further by tuning them under a training objective that en- courages their gradient-based explanations to be consistent with human-provided region-based annotations from visu- ally grounded data when those are available. A picture of a cathedral next to a park A picture of a cathedral next to a parkA picture of a cathedral next to a parkInput Image + TextRegular V-L Model Explanation Human Explanationw/ Attention Mask Consistency (Ours)A picture of a cathedral next to a parkFigure 1. Gradient-based methods can generate heatmaps that ex- plain the match between images and text for a Vision-language model (VLM). Our work aims to improve their ability to produce visual groundings by directly optimizing their gradient-based ex- planations so that they are consistent with human annotations pro- vided for a reduced set of images. Vision-language transformers extend the success of masked language modeling (MLM) to multi-modal prob- lems. In vision-language transformers, objectives such as image-text matching (ITM), and image-text contrastive losses (ITC) are used in addition to MLM to exploit com- monalities between images and text [6, 17, 18, 22]. We fur- ther extend these objectives to include our proposed Atten- tion Mask Consistency (AMC) objective. Our formulation is based on the observation that gradient-based explanation maps obtained using methods such as GradCAM [30], can be used to explain the image-text matching of a VLM. Our AMC objective explicitly optimizes these explanations dur- ing training so that they are consistent with region annota- tions. Figure 1 illustrates an example input image and text This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 19165 an empty streetGradCAM(/uni2207y)/uni03D5v/uni03D5ty/uni03D5fM:/uni2112amcV:T: {} {}maximizeminimizeFigure 2. Overview of our method. Among other objectives, standard vision-language models are trained to produce a matching score ygiven an input image-text pair (V, T ). For inputs containing an extra level of supervision in the form of region annotations (e.g. a triplet (V, T, M )), where M is a binary mask indicating the regions annotated by a human, we optimize the GradCAM [30] gradient-based explanations of the model so that the produced explanations are consistent with region annotations using Lamcby maximizing the energy in the heatmap that falls inside the region annotation and minimizing what falls outside. We accomplish this through soft margin losses as described in Sec. 3.2. pair along with a gradient-based explanation obtained from a VLM model, a region annotation provided by a human, and an improved gradient-based explanation after the VLM model was tuned under our proposed objective. Our work builds particularly upon the ALBEF model [17] which incorporates a vision-language model architecture based on transformers [36] and has already demonstrated off-the-shelf grounding capabilities using GradCAM. Gradient-based explanations in the form of heatmaps have been used extensively to explain the areas of the input images that most impact an output value of a model. In our formulation we actively leverage these heatmaps by designing a loss function that encourages most of the energy in the heatmaps to fall within the areas of the input image that most align with human provided region annotations. Figure 2 shows a detailed overview of our method and objective function. Given an input image and text pair, our goal is to maximize a soft margin between the energy of the heatmap inside the region annotation and the energy of the heatmap outside the region annotation. A soft-margin is important since typical human region annotations in the form of boxes do not exactly outline objects of different shapes, and in many cases models should still be able to ground an input text with multiple regions across the image. We compare AMC extensively against other methods that use the same level of supervision but instead use an object detector such as Faster-RCNN [10, 11, 17, 23]. Our method obtains state-of-the-art pointing game accuracy on both Flickr30k and RefCOCO+. Our contributions can be summarized as follows: (1) We introduce a new training objective, AMC, which is effective, simple to implement and can handle multiple types of region annotations, (2) Weshow that AMC can improve the grounding capabilities of an existing vision-language model – ALBEF, and (3) the resulting model is state-of-the-art in two benchmarks for phrase grounding and referring expression comprehension.
Yan_GCFAgg_Global_and_Cross-View_Feature_Aggregation_for_Multi-View_Clustering_CVPR_2023	Abstract Multi-view clustering can partition data samples into their categories by learning a consensus representation in unsupervised way and has received more and more atten- tion in recent years. However, most existing deep clustering methods learn consensus representation or view-specific representations from multiple views via view-wise aggre- gation way, where they ignore structure relationship of all samples. In this paper, we propose a novel multi-view clus- tering network to address these problems, called Global and Cross-view Feature Aggregation for Multi-View Clus- tering (GCFAggMVC). Specifically, the consensus data pre- sentation from multiple views is obtained via cross-sample and cross-view feature aggregation, which fully explores the complementary of similar samples. Moreover, we align the consensus representation and the view-specific representa- tion by the structure-guided contrastive learning module, which makes the view-specific representations from differ- ent samples with high structure relationship similar. The proposed module is a flexible multi-view data representa- tion module, which can be also embedded to the incomplete multi-view data clustering task via plugging our module into other frameworks. Extensive experiments show that the proposed method achieves excellent performance in both complete multi-view data clustering tasks and incomplete multi-view data clustering tasks.	1. Introduction With the rapid development of informatization, data is often collected by various social media or views. For in- stance, a 3D object can be described from different angles; a news event is reported from different sources; and an im- age can be characterized by different types of feature sets, e.g., SIFT, LBP, and HoG. Such an instance object, which is *Corresponding author([email protected])described from multiple views, is referred to as multi-view data. Multi-view clustering (MVC) [6], i.e., unsupervisedly fusing the multi-view data to aid differentiate crucial group- ing, is a fundamental task in the fields of data mining, pat- tern recognition, etc, but it remains a challenging problem. Traditional multi-view clustering methods [7] include matrix decomposition methods, graph-based multi-view methods, and subspace-based multi-view methods. The goal of these methods is to obtain a high-quality consensus graph or subspace self-representation matrix by various reg- ularization constraints in order to improve the performance of clustering. However, most of them directly operate on the original multiview features or specified kernel features, which usually include noises and redundancy information during the collection or kernel space selection processes, moreover harmful to the clustering tasks. Deep neural networks have demonstrated excellent per- formance in data feature representation for many vision tasks. Deep clustering methods also draw more attention to researchers [1, 9, 40, 50–52]. These methods efficiently learn the feature presentation of each view using a view- specific encoder network, and fuse these learnt representa- tions from all views to obtain a consensus representation that can be divided into different categories by a cluster- ing module. To reduce the influence of view-private in- formation on clustering, these methods designed different alignment models. For example, some methods align the representation distributions or label distributions from dif- ferent views by KL divergence [14]. They might be hard to distinguish between clusters, since a category from one view might be aligned with a different category in another view. Some methods align the representation from different views by contrastive learning. Despite these models have achieved significant improvement in MVC task, the fol- lowing issues still exist: 1) Almost all existing deep MVC methods (such as [38,40,49]) are based on view-wise fusion models, such as weighted-sum fusion of all views or con- catenating fusion of all views, which makes it difficult to This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 19863 obtain discriminative consensus representations from mul- tiple views, since a view or several views of a sample might contain too much noise or be missing
Zhang_Transforming_Radiance_Field_With_Lipschitz_Network_for_Photorealistic_3D_Scene_CVPR_2023	Abstract Recent advances in 3D scene representation and novel view synthesis have witnessed the rise of Neural Radiance Fields (NeRFs). Nevertheless, it is not trivial to exploit NeRF for the photorealistic 3D scene stylization task, which aims to generate visually consistent and photorealistic styl- ized scenes from novel views. Simply coupling NeRF with photorealistic style transfer (PST) will result in cross-view inconsistency and degradation of stylized view syntheses. Through a thorough analysis, we demonstrate that this non- trivial task can be simplified in a new light: When trans- forming the appearance representation of a pre-trained NeRF with Lipschitz mapping, the consistency and pho- torealism across source views will be seamlessly encoded into the syntheses. That motivates us to build a concise and flexible learning framework namely LipRF , which upgrades arbitrary 2D PST methods with Lipschitz mapping tailored for the 3D scene. Technically, LipRF first pre-trains a ra- diance field to reconstruct the 3D scene, and then emulates the style on each view by 2D PST as the prior to learn a Lip- schitz network to stylize the pre-trained appearance. In view of that Lipschitz condition highly impacts the expressivity of the neural network, we devise an adaptive regularization to balance the reconstruction and stylization. A gradual gra- dient aggregation strategy is further introduced to optimize LipRF in a cost-efficient manner. We conduct extensive ex- periments to show the high quality and robust performance of LipRF on both photorealistic 3D stylization and object appearance editing.	1. Introduction Photorealistic style transfer (PST) [52] is one of the im- portant tasks for visual content creation, which aims to au- tomatically apply the color style of a reference image to an- *Corresponding authorother input ( e.g., image [38] or video [65]). In this task, the stylized result is required to look like a camera shot and preserve the input structure ( e.g., edges and regions). Benefiting from the launch of deep learning, a series of so- phisticated deep PST methods [19,30,38,59,65] have been developed for practical usage. Recent progress in 3D scene representation has featured Neural Radiance Field [43, 67] (NeRF) with efficient training and high-quality view syn- thesis. This inspires us to go one step further to explore a more challenging task of photorealistic 3D scene styl- ization, which is to generate visually consistent and pho- torealistic stylized syntheses from arbitrary views. Such a task enables an automatic modification of 3D scene appear- ance with different lighting, time of day, weather, or other effects, thereby enhancing user experience and stimulating emotions for virtual reality [57]. Nevertheless, it is not trivial to build an effective frame- work for photorealistic 3D scene stylization. The diffi- culty mainly originates from the fact that there is no valid photorealistic style loss tailored for training NeRF. In gen- eral, the image-based PST is commonly tackled via either the neural style transfer [11] combined with complicated post-processing [30, 38, 41], or particular network struc- tures [29, 61, 65]. However, none of them can be directly applied to the learning of NeRF. As shown in Figure 1, sim- ply employing the state-of-the-art 2D PST on each view might result in noise, disharmony and even inconsistency across views, since the PST methods rely on the size or ob- ject masks of the inputs. Such downsides will be further amplified after reconstructing the 3D scene with NeRF. To alleviate these limitations, we start with a basic un- derstanding of this task: Though preserving the photoreal- ism and consistency seems to be different in the context of 2D images, they do have the same essence when moving to 3D volume rendering [40]. From this standpoint, the task is simplified as a problem to regulate the volume rendering variance of the radiance field before and after stylization. According to the studies of color mapping [48,51,52], some This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 20712 Scene &Style Radiance field Radiance field w/ Lipschitz MLP 2D Photorealistc style transfer Figure 1. Illustrations of different strategies for photorealistic 3D scene stylization. The 2D PST method [9] generates disharmonious orange color on the stem and petaline edge. The regions bounded by the dotted ellipses are inconsistent due to the white spot in the first view. When employing a radiance field to reconstruct the results of 2D PST method, the disharmony and inconsistency are still retained. Our LipRF successfully eliminates these downsides, and renders high-quality stylized view syntheses. specific linear mappings of image pixels can nicely preserve the image structures with photorealistic effect. Motivated by this, we theoretically demonstrate that a simple yet ef- fective design of Lipschitz-constrained linear mapping over appearance representation can elegantly control the volume rendering variance. Furthermore, we prove that replac- ing linear mapping with a Lipschitz multilayer perceptron (MLP) also holds these properties under extra assumptions which can be relaxed in practice. Such a way completely eliminates the drawbacks of 2D PST when transforming the radiance field with a Lipschitz MLP (see Figure 1). In a nutshell, our analysis verifies that Lipschitz MLP can be in- terpreted as an implicit regularization to safeguard the 3D photography of stylized scenes. By consolidating the idea of transforming the radiance field with the Lipschitz network, we propose a novel NeRF- based architecture (namely LipRF) for photorealistic 3D scene stylization. Technically, LipRF contains two stages: 1) training a radiance field to reconstruct the source 3D scene; 2) learning a Lipschitz network to transform the pre- trained appearance representation to the stylized 3D scene with the guidance of style emulation on each view by ar- bitrary 2D PST. We adopt the Plenoxels [67] as the base radiance field due to its advanced reconstruction quality and compressed appearance representation by spheric har- monics. Considering that the Lipschitz condition greatly impacts the expressivity of neural networks, we design an adaptive regularization based on spectral normalization [44] to allow a mild relaxation of the Lipschitz constant in each linear layer. Finally, we capitalize on gradual gradient ag- gregation to optimize LipRF in a cost-efficient fashion. In summary, we have made the following contributions: (I) We present a thorough and insightful analysis of photo- realistic 3D scene stylization on the basis of volume render- ing and Lipschitz transformation. ( II) We build a concise and flexible framework (LipRF) for photorealistic 3D scene stylization by novelly transforming the appearance repre- sentation of a pre-trained NeRF with the Lipschitz Network. (III) Under the Lipschitz condition, we design adaptive reg- ularization and gradual gradient aggregation to seek a bettertrade-off among the reconstruction, stylization quality, and computational cost. We evaluate LipRF on both photoreal- istic 3D stylization and object appearance editing tasks to validate the effectiveness of our proposal.
Yu_Semi-Supervised_Domain_Adaptation_With_Source_Label_Adaptation_CVPR_2023	Abstract Semi-Supervised Domain Adaptation (SSDA) involves learning to classify unseen target data with a few labeled and lots of unlabeled target data, along with many labeled source data from a related domain. Current SSDA ap- proaches usually aim at aligning the target data to the la- beled source data with feature space mapping and pseudo- label assignments. Nevertheless, such a source-oriented model can sometimes align the target data to source data of the wrong classes, degrading the classification per- formance. This paper presents a novel source-adaptive paradigm that adapts the source data to match the tar- get data. Our key idea is to view the source data as a noisily-labeled version of the ideal target data. Then, we propose an SSDA model that cleans up the label noise dy- namically with the help of a robust cleaner component de- signed from the target perspective. Since the paradigm is very different from the core ideas behind existing SSDA ap- proaches, our proposed model can be easily coupled with them to improve their performance. Empirical results on two state-of-the-art SSDA approaches demonstrate that the proposed model effectively cleans up the noise within the source labels and exhibits superior performance over those approaches across benchmark datasets. Our code is avail- able at https://github.com/chu0802/SLA .	1. Introduction Domain Adaptation (DA) focuses on a general machine learning scenario where training and test data may originate from two related but distinct domains: the source domain and the target domain. Many works have extensively stud- ied unsupervised DA (UDA), where labels in the target do- main cannot be accessed, from both theoretical [2, 19, 36] and algorithmic [5, 8, 15, 16, 22, 37] perspectives. Recently, Semi-Supervised Domain Adaptation (SSDA), another DA setting that allows access to a few target labels, has received more research attention because it is a simple yet realistic setting for application needs. The most na ¨ıve strategy for SSDA, commonly known as Figure 1. Top. Training the model with the original source labels might lead to the misalignment of the target data. Bottom. After cleaning up the noisy source labels with our SLA framework, the target data can be aligned with the correct classes. S+T [21, 33], aims to train a model using the source data and labeled target data with a standard cross entropy loss. This strategy often suffers from a well-known domain shift issue, which stems from the gap between different data dis- tributions. To address this issue, many state-of-the-art al- gorithms attempt to explore better use of the unlabeled tar- get data so that the target distribution can be aligned with the source distribution. Recently, several Semi-Supervised Learning (SSL) algorithms have been applied for SSDA [12, 21, 30] to regularize the unlabeled data, such as en- tropy minimization [6], pseudo-labeling [11,24] and consis- tency regularization [1, 24]. These classic source-oriented strategies have prevailed for a long time. However, these algorithms typically require the target data to closely match some semantically similar source data in the feature space. Therefore, if the S+T space has been misaligned, it can be challenging to recover from the misalignment, as illustrated in Figure 1. We take a deeper look into a specific example from the Office-Home dataset [27] to confirm the abovementioned is- sue. Figure 2 visualizes the feature space trained by S+T us- ing t-SNE [3]. We observed that the misalignment between This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 24100 Figure 2. Feature visualizations with t-SNE for an example of the misalignment on the Office-Home A→C dataset with ResNet34. The model is trained by S+T. Left: 0-th iteration. Right : 5000-th iteration . We observe that the misalignment has already happened at a very early stage. Guided by source labels and a few target labels, a portion of the target data from the 59th class misaligns with the source data from the 7th class. the source and the target data has happened at a very early stage. For instance, in the beginning, a portion of the target data from the 59th class is close to the source data from the 7th class. Since we only have access to source labels and a few target labels, without proper guidance from enough target labels, such misalignment becomes more severe after being trained by S+T. Table 1 shows the partial confusion matrix of S+T. Roughly 40% of the target data in the 59th class is mispredicted to the 7th class, and only around 20% of the data is classified correctly. From the case study above, we argue that relying on the source labels like S+T can misguide the model to learn wrong classes for some target data. That is, source labels can be viewed as a noisy version of the ideal labels for target classification. Based on the argument, the setting of SSDA is more like a Noisy Label Learning (NLL) problem, with a massive amount of noisy labels (source labels) and a small number of clean labels (target labels). Learning with noisy labels is a widely studied machine learning problem. A popular solution is to clean up the noisy labels with the help of another model, which can also be referred to as label correction [28]. To approach Do- main Adaptation as an NLL problem, we borrow the idea from label correction and propose a Source Label Adapta- tion (SLA) framework, as shown in Figure 1. We construct a label adaptation component that provides the view from the target data and dynamically cleans up the noisy source labels at each iteration. Unlike other earlier works that study how to leverage the unlabeled target data, we mainly inves- tigate how to train the source data with the adapted labels to better suit the ideal target space. This source-adaptive paradigm is entirely orthogonal to the core ideas behind ex- isting SSDA algorithms. Thus, we can combine our frame- work with those algorithms to get superior results. We sum- marize our contributions as follows. • We argue that the classic source-oriented methodsTrue\Pred Class 7 Class 59 Class 41 Others Class 59 38.5% 19.8% 13.5% 28.2% Table 1. A partial confusion matrix of S+T on the 3-shot Office- Home A→C dataset with ResNet34. might still suffer from the biased feature space de- rived from S+T. To escape this predicament, we pro- pose adapting the source data to the target space by modifying the original source labels. • We address DA as a particular case of NLL problems and present a novel source-adaptive paradigm. Our SLA framework can be easily coupled with other ex- isting algorithms to boost their performance. • We demonstrate the usefulness of our proposed SLA framework when coupled with state-of-the-art SSDA algorithms. The framework significantly improved ex- isting algorithms on two major benchmarks, inspiring a new direction for solving DA problems.
Zhang_Gradient_Norm_Aware_Minimization_Seeks_First-Order_Flatness_and_Improves_Generalization_CVPR_2023	Abstract Recently, flat minima are proven to be effective for im- proving generalization and sharpness-aware minimization (SAM) achieves state-of-the-art performance. Yet the cur- rent definition of flatness discussed in SAM and its follow- ups are limited to the zeroth-order flatness (i.e., the worst- case loss within a perturbation radius). We show that the zeroth-order flatness can be insufficient to discriminate min- ima with low generalization error from those with high gen- eralization error both when there is a single minimum or multiple minima within the given perturbation radius. Thus we present first-order flatness, a stronger measure of flat- ness focusing on the maximal gradient norm within a per- turbation radius which bounds both the maximal eigenvalue of Hessian at local minima and the regularization function of SAM. We also present a novel training procedure named Gradient norm Aware Minimization (GAM) to seek minima with uniformly small curvature across all directions. Ex- perimental results show that GAM improves the generaliza- tion of models trained with current optimizers such as SGD and AdamW on various datasets and networks. Further- more, we show that GAM can help SAM find flatter minima and achieve better generalization. The code is available at https://github.com/xxgege/GAM .	1. Introduction Current neural networks have achieved promising results in a wide range of fields [36, 53, 55, 67, 73–75, 78], yet they are typically heavily over-parameterized [2, 3]. Such heavy overparameterization leads to severe overfitting and poor generalization to unseen data when the model is learned simply with common loss functions (e.g., cross-entropy) [27]. Thus effective training algorithms are required to limit the negative effects of overfitting training data and find gen- †Equal contribution, *Corresponding author ZOFZOF FOFFOF ρ ρTraining Testing(a) ZOF ZOF FOFFOF ρ ρTraining Testing (b) Figure 1. The comparison of the zeroth-order flatness (ZOF) and first-order flatness (FOF). Given a perturbation radius ρ, ZOF can fail to indicate generalization error both when there are multi- ple minima (1a) and a single minimum (1b) in the radius while FOF remains discriminative. The height of blue rectangles in curly brackets is the value of ZOF and the height of gray trian- gles (which indicates the slope) is the value of FOF. In Figure 1a, when ρis large and enough to cover multiple minima, ZOF can not measure the fluctuation frequency while FOF prefers the flat- ter valley which has a smaller gradient norm. When ρis small and covers only a single minimum, the maximum loss in ρcan be misleading as it can be misaligned with the uptrend of loss. As shown in Figure 1b, ZOF prefers the valley on the right, which has a larger generalization error (the orange dotted line), while FOF prefers the left one. eralizable solutions. Many studies try to improve model generalization by modifying the training procedure, such as batch normal- ization [26], dropout [24], and data augmentation [13, 68, 72]. Especially, some works discuss the connection be- tween the geometry of the loss landscape and general- ization [19, 22, 27]. A branch of effective approaches, sharpness-Aware Minimization (SAM) [19] and its variants [16,17,34,43,48,77], minimizes the worst-case loss within a perturbation radius, which we call zeroth-order flatness. It is proven that optimizing the zeroth-order flatness leads to lower generalization error and achieves state-of-the-art per- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 20247 formance on various image classification tasks [19, 39, 80]. Optimizing the worst case, however, relies on a reason- able choice of perturbation radius ρ. As a prefixed hyper- parameter in SAM or a hyperparameter under parameter re- scaling in its variants, such as ASAM [39], ρcan not al- ways be a perfect choice in the whole training process. We show that the zeroth-order flatness may fail to indicate the generalization error with a given ρ. As in Figure 1a, when ρcovers multiple minima, the zeroth-order flatness (SAM) can not measure the fluctuation frequency. When there is a single minimum within ρ, as in Figure 1b the observation radius is limited and the maximum loss in ρcan be mis- aligned with the uptrend of loss. So zeroth-order flatness can be misleading and the knowledge of loss gradient is re- quired for generalization error minimization. To address this problem, we introduce first-order flat- ness, which controls the maximum gradient norm in the neighborhood of minima. We show that the first-order flat- ness is stronger than the zeroth-order flatness as the loss intensity of the loss fluctuation can be bounded by the max- imum gradient. When the perturbation radius covers mul- tiple minima, which we show is quite common in prac- tice, the first-order flatness discriminates more drastic jit- ters from real flat valleys, as in Figure 1a. When the per- turbation radius is small and covers only one minimum, the first-order flatness demonstrates the trend of loss gradient and can help indicate generalization error. We further show that the first-order flatness directly controls the maximal eigenvalue of Hessian of the training loss, which is a proper sharpness/flatness measure indicating the loss uptrend un- der an adversarial perturbation to the weights [31–33]. To optimize the first-order flatness in deep model training, we propose Gradient norm Aware Minimization (GAM), which approximates the maximum gradient norm with stochastic gradient ascent and Hessian-vector products to avoid the materialization of the Hessian matrix. We summarize our contributions as follows. • We present first-order flatness, which measures the largest gradient norm in the neighborhood of minima. We show that the first-order flatness is stronger than current zeroth- order flatness and it controls the maximum eigenvalue of Hessian. • We propose a novel training procedure, GAM, to simulta- neously optimize prediction loss and first-order flatness. We analyze the generalization error and the convergence of GAM. • We empirically show that GAM considerably improves model generalization when combined with current opti- mizers such as SGD and AdamW across a wide range of datasets and networks. We show that GAM further im- proves the generalization of models trained with SAM. • We empirically validate that GAM indeed finds flatter op- tima with lower Hessian spectra.
Yang_TINC_Tree-Structured_Implicit_Neural_Compression_CVPR_2023	Abstract Implicit neural representation (INR) can describe the target scenes with high fidelity using a small number of param- eters, and is emerging as a promising data compression technique. However, limited spectrum coverage is intrinsic to INR, and it is non-trivial to remove redundancy in diverse complex data effectively. Preliminary studies can only exploit either global or local correlation in the target data and thus of limited performance. In this paper, we propose a Tree-structured Implicit Neural Compression (TINC) to conduct compact representation for local regions and extract the shared features of these local representations in a hierarchical manner. Specifically, we use Multi-Layer Perceptrons (MLPs) to fit the partitioned local regions, and these MLPs are organized in tree structure to share pa- rameters according to the spatial distance. The parameter sharing scheme not only ensures the continuity between adjacent regions, but also jointly removes the local and non-local redundancy. Extensive experiments show that TINC improves the compression fidelity of INR, and has shown impressive compression capabilities over commer- cial tools and other deep learning based methods. Besides, the approach is of high flexibility and can be tailored for different data and parameter settings. The source code can be found at https://github.com/RichealYoung/TINC.	1. Introduction In the big data era, there exist massive and continuously growing amount of visual data from different fields, from surveillance, entertainment, to biology and medical diagno- sis. The urgent need for efficient data storage, sharing and transmission all requires effective compression techniques. Although image and video compression have been studied for decades and there are a number of widely used commer- cial tools, the compression of medical and biological data This work was done in collaboration with Tingxiong Xiao and Yux- iao Cheng, under supervision of Prof. Jinli Suo and Prof. Qionghai Dai affiliated with Department of Automation, Tsinghua Univ.are not readily available. Implicit neural representation (INR) is becoming widely-used for scene rendering [19, 21, 23, 31, 44], shape estimation [8, 11, 25, 29], and dynamics modeling [15, 26, 28]. Due to its powerful representation capability, INR can describe nature scenes at high fidelity with a much smaller number of parameters than raw discrete grid representation, and thus serves as a promising data compression technique. In spite of the big potential, INR based compression is quite limited confronted with large sized data [20], since INR is intrinsically of limited spectrum coverage and can- not envelop the spectrum of the target data, as analyzed in [41]. Two pioneering works using INR for data com- pression, including NeRV [7] and SCI [41], have attempted to handle this issue in their respective ways. Specifically, NeRV introduces the convolution operation into INR, which can reduce the required number of parameters using the weight sharing mechanism. However, convolution is spa- tially invariant and thus limits NeRV’s representation accu- racy on complex data with spatial varying feature distribu- tion. Differently, to obtain high fidelity on complex data, SCI adopts divide-and-conquer strategy and partitions the data into blocks within INR’s concentrated spectrum en- velop. This improves the local fidelity, but cannot remove non-local redundancies for higher compression ratio and tend to cause blocking artifacts. To compress large and complex data with INR, in this pa- per we propose to build a tree-structured Multi-Layer Per- ceptrons (MLPs), which consists of a set of INRs to rep- resent local regions in a compact manner and organizes them under a hierarchical architecture for parameter sharing and higher compression ratio. Specifically, we first draw on the idea of ensemble learning [12] and use a divide- and-conquer strategy [20, 41] to compress different regions with multiple MLPs separately. Then we incorporate these MLPs with a tree structure to extract their shared param- eters hierarchically, following the rule that spatially closer regions are of higher similarity and share more parameters. Such a joint representation strategy can remove the redun- dancy both within each local region and among non-local This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 18517 regions, and also ensures the continuity between adjacent regions. The scheme of the proposed network is illustrated in Fig. 1, and we name our approach TINC (Tree-structured Implicit Neural Compression). Using the massive and diverse biomedical data, we con- duct extensive experiments to validate that TINC greatly improves the capability of INR and even outperforms the commercial compression tools (H.264 and HEVC) under high compression ratios. The proposed TINC is also a gen- eral framework, which can be flexibly adapted to diverse data and varying settings.
Zhang_Unlearnable_Clusters_Towards_Label-Agnostic_Unlearnable_Examples_CVPR_2023	Abstract There is a growing interest in developing unlearnable examples (UEs) against visual privacy leaks on the Internet. UEs are training samples added with invisible but unlearn- able noise, which have been found can prevent unauthorized training of machine learning models. UEs typically are gen- erated via a bilevel optimization framework with a surrogate model to remove (minimize) errors from the original sam- ples, and then applied to protect the data against unknown target models. However, existing UE generation methods all rely on an ideal assumption called label-consistency , where the hackers and protectors are assumed to hold the same label for a given sample. In this work, we propose and promote a more practical label-agnostic setting, where the hackers may exploit the protected data quite differently from the protectors. E.g., a m-class unlearnable dataset held by the protector may be exploited by the hacker as a n-class dataset. Existing UE generation methods are ren- dered ineffective in this challenging setting. To tackle this challenge, we present a novel technique called Unlearn- able Clusters (UCs) to generate label-agnostic unlearn- able examples with cluster-wise perturbations. Furthermore, we propose to leverage Vision-and-Language Pre-trained Models (VLPMs) like CLIP as the surrogate model to im- prove the transferability of the crafted UCs to diverse do- mains. We empirically verify the effectiveness of our pro- posed approach under a variety of settings with different datasets, target models, and even commercial platforms Mi- crosoft Azure and Baidu PaddlePaddle . Code is avail- able at https://github.com/jiamingzhang94/ Unlearnable-Clusters .	1. Introduction While the huge amount of “free” data available on the Internet has been key to the success of deep learning and computer vision, this has also raised public concerns on the unauthorized exploitation of personal data uploaded to the *This work is done when the author interned at Peng Cheng Lab. †Corresponding authors. Unconscious and unauthorized collection for training Label -consistency Label -agnosticOriginal images 𝒙𝒙 Unlearnable image 𝒙𝒙′ Hacker Protector 𝒎𝒎-class surrogate model 𝒇𝒇𝒔𝒔𝒎𝒎 =car, dog, cat, airplane, etc. ... 𝒎𝒎-class target model 𝒇𝒇𝒕𝒕𝒎𝒎 car, dog, cat, airplane, etc.= 𝒏𝒏-class target model 𝒇𝒇𝒕𝒕𝒏𝒏 =animal, vehicle, etc. ... HackerFigure 1. An illustration of two different data protection assump- tions: label-consistency vs.label-agnostic , where the hacker ex- ploits the protected data in different manners. Internet to train commercial or even malicious models [16]. For example, a company named Clearview AI has been found to have scraped billions of personal images from Face- book, YouTube, Venmo and millions of other websites to construct a commercial facial recognition application [44]. This has motivated the proposal of Unlearnable Examples (UEs) [17] to make data unlearnable (or unusable) to ma- chine learning models/services. Similar techniques are also known as availability attacks [2, 41] or indiscriminate poi- soning attacks [14] in the literature. These techniques allow users to actively adding protective noise into their private data to avoid unauthorized exploitation, rather than putting our trust into the hands of large corporations. The original UE generation method generates error- minimizing noise via a bilevel min-min optimization frame- work with a surrogate model [17]. The noise can then be added to samples in a training set in either a sample-wise or class-wise manner to make the entire dataset unlearnable to different DNNs. It has been found that this method can- not survive adversarial training, which has been addressed by a recent method [11]. In this work, we identify one common assumption made by existing UE methods: label- consistency , where the hackers will exploit the protected dataset in the same way as the protector including the labels. This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 3984 This means that, for the same image, the hacker and protec- tor hold the same label. We argue that this assumption is too ideal, and it is possible that the hackers will collect the protected (unlearnable) samples into a dataset for a different task and label the dataset into different number of classes. As illustrated in Figure 1, an image can be labelled with different annotated labels (cat or animal), showing that a m- class (e.g., 10-class) unlearnable dataset may be exploited by the hacker as a n-class (e.g., 5-class or 20-class) dataset depending on its actual needs. We term this more generic assumption as label-agnostic and propose a novel method Unlearnable Clusters (UCs) to generate more effective and transferable unlearnable examples under this harsh setting. In Figure 2 (a), we show that this more generic label- agnostic setting poses a unique transferability challenge for the noise generated by existing methods like Error- Minimizing Noise (EMinN) [17], Adversarial Poisoning (AdvPoison) [10], Synthetic Perturbations (SynPer) [41] and DeepConfuse [9]. This indicates that the protective noise generated by these methods are label-dependent and are ren- dered ineffective when presented with different number of classes. As such, we need more fundamental approaches to make a dataset unlearnable regardless of the annotations. To this end, we start by analyzing the working mechanism of UEs generated by EMinN, AdvPoison as they are very representative under the label-consistency setting. Through a set of visual analyses, we find that the main reason why they could break supervised learners is that the generated noise tends to disrupts the distributional uniformity and dis- crepancy in the deep representation space. Uniformity refers to the property that the manifold of UEs in the deep rep- resentation space does not deviate much from that of the clean examples, while discrepancy refers to the property that examples belonging to the same class are richly diverse in the representation space. Inspired by the above obser- vation, we propose a novel approach called Unlearnable Clusters (UCs) to generate label-agnostic UEs using cluster- wise (rather than class-wise) perturbations. This allows us to achieve a simultaneous disruption of the uniformity and discrepancy without knowing the label information. Arguably, the choose of a proper surrogate model also plays an important role in generating effective UEs. Previ- ous methods generate UEs by directly attacking a surrogate model and then transfer the generated UEs to fight against a diverse set of target models [10, 17]. This may be easily achievable under the label-consistency setting, but may fail badly under the label-agnostic setting. However, even un- der the label-consistency setting, few works have studied the impact of the surrogate model to the final unlearnable performance. To generate effective, and more importantly, transferable UEs under the label-agnostic setting, we need to explore more generic surrogate model selection strategies, especially those that can be tailored to a wider range of un-known target models. Intuitively, the surrogate model should be a classification DNN that contains as many classes as possible so as to facilitate the recognition and protection of billions of images on the Internet. In this paper, we propose to leverage the large-scale Vision-and-Language Pre-trained Models (VLPMs) [22,23,30] like CLIP [30] as the surrogate model. Pre-trained on over 400 million text-to-image pairs, CLIP has the power to extract the representation of extremely diverse semantics. Meanwhile, VLPMs are pre-trained with a textual description rather than a one-hot label to align with the image, making them less overfit to the actual class “la- bels”. In this work, we leverage the image encoder of CLIP to extract the embeddings of the input images and then use the embeddings to generate more transferable UCs. We evaluate our UC approach with different backbones and datasets, all in a black-box setting (the protector does not know the attacker’s network architecture or the class labels). Cluster-wise unlearnable noise can also prevent un- supervised exploitation against contrastive learning to certain extent, proving its superiority to existing UEs. We also com- pare UC with existing UE methods against two commercial machine learning platforms: Microsoft Azure1and Baidu PaddlePaddle2. To the best of our knowledge, this is the first physical-world attack to commercial APIs in this line of work. Our main contributions are summarized as follows: •We promote a more generic data protection assumption called label-agnostic , which allows the hackers to ex- ploit the protected dataset differently (in terms of the annotated class labels) as the protector. This opens up a more practical and challenging setting against unau- thorized training of machine learning models. •We reveal the working mechanism of existing UE gener- ation methods: they all disrupt the distributional unifor- mity and discrepancy in the deep representation space. •We propose a novel approach called Unlearnable Clus- ters(UCs) to generate label-agnostic UEs with cluster- wise perturbations without knowing the label informa- tion. We also leverage VLPMs like CLIP as the surro- gate model to craft more transferable UCs. •We empirically verify the effectiveness of our proposed approach with different backbones on different datasets. We also show its effectiveness in protecting private data against commercial machine learning platforms Azure andPaddlePaddle .
Yang_ContraNeRF_Generalizable_Neural_Radiance_Fields_for_Synthetic-to-Real_Novel_View_Synthesis_CVPR_2023	Abstract Although many recent works have investigated general- izable NeRF-based novel view synthesis for unseen scenes, they seldom consider the synthetic-to-real generalization, which is desired in many practical applications. In this work, we first investigate the effects of synthetic data in synthetic-to-real novel view synthesis and surprisingly ob- serve that models trained with synthetic data tend to pro- duce sharper but less accurate volume densities. For pix- els where the volume densities are correct, fine-grained de- tails will be obtained. Otherwise, severe artifacts will be produced. To maintain the advantages of using synthetic data while avoiding its negative effects, we propose to intro- duce geometry-aware contrastive learning to learn multi- view consistent features with geometric constraints. Mean- while, we adopt cross-view attention to further enhance the geometry perception of features by querying features across input views. Experiments demonstrate that under the synthetic-to-real setting, our method can render images with higher quality and better fine-grained details, outper- forming existing generalizable novel view synthesis meth- ods in terms of PSNR, SSIM, and LPIPS. When trained on real data, our method also achieves state-of-the-art results. https://haoy945.github.io/contranerf/	1. Introduction Novel view synthesis is a classical problem in computer vision, which aims to produce photo-realistic images for un- seen viewpoints [2,5,10,36,40]. Recently, Neural Radiance Fields (NeRF) [25] proposes to achieve novel view synthe- sis through continuous scene modeling through a neural net- *Joint last authorship.work, which quickly attracts widespread attention due to its surprising results. However, the vanilla NeRF is actu- ally designed to fit the continuous 5D radiance field of a given scene, which often fails to generalize to new scenes and datasets. How to improve the generalization ability of neural scene representation is a challenging problem. Recent works, such as pixelNeRF [46], IBRNet [37], MVSNeRF [4] and GeoNeRF [19], investigate how to achieve generalizable novel view synthesis based on neu- ral radiance fields. However, these works mainly focus on the generalization of NeRF to unseen scenes and seldom consider the synthetic-to-real generalization, i.e., training NeRF with synthetic data while testing it on real data. On the other hand, synthetic-to-real novel view synthesis is de- sired in many practical applications where the collection of dense view 3D data is expensive (e.g., autonomous driv- ing, robotics, and unmanned aerial vehicle [34]). Although some works directly use synthetic data such as Google Scanned Objects [14] in model training, they usually over- look the domain gaps between the synthetic and real data as well as possible negative effects of using synthetic data. In 2D computer vision, it is common sense that synthetic training data usually hurts the model’s generalization abil- ity to real-world applications [1, 8, 44]. Will synthetic data be effective in novel view synthesis? In this work, we first investigate the effectiveness of syn- thetic data in NeRF’s training via extensive experiments. Specifically, we train generalizable NeRF models using a synthetic dataset of indoor scenes called 3D-FRONT [15], and test the models on a real indoor dataset called Scan- Net [9]. Surprisingly, we observe that the use of synthetic data tends to result in more artifacts on one hand but bet- ter fine-grained details on the other hand (see Fig.1 and Sec.3.2 for more details). Moreover, we observe that mod- els trained on synthetic data tend to predict sharper but less This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 16508 (1)(2)(a) Ground truth (b) Pred (ScanNet) (c) Pred (3D-FRONT) (2)(1) (d) Density (ScanNet) (1)(2) (e) Density (3D-FRONT) Figure 1. Fig.1a is the ground truth of the target image to be rendered. Fig.1b and Fig.1c are the rendered images when models are trained on ScanNet [9] and 3D-FRONT [15], respectively. Compare to Fig.1b, Fig.1c is more detailed (see purple box) but has more artifacts (see pink box). Fig.1d and Fig.1e further show the volume density (the redder the color, the higher the density) along the epipolar line projected from the orange points in Fig.4a to the source view. The model trained on 3D-FRONT prefers to predict the volume density with a sharper distribution, but sometimes the predicted volume density is not accurate (see line 1 in Fig.1e), resulting in severe artifacts. (a) Deviation. (b) Error. Figure 2. Deviation and error of predicted depth when trained with synthetic and real data, respectively. We count the devi- ation and error of the predicted depth for each pixel in the test dataset and plot them as the histogram. The depth is calculated by aggregating depth of the sampled points along the rendering ray, similar to the process of color rendering. Compared to the model trained with the real data, the model trained with the synthetic data tends to predict depths with small deviations but large errors, i.e., density distributions that are sharper but less geometrically accu- rate. accurate volume densities (see Fig.2). In this case, better fine-grained details can be obtained once the prediction of geometry (i.e., volume density) is correct, while severe arti- facts will be produced otherwise. This motivates us to con- sider one effective way to generalize from synthetic data to real data in a geometry-aware manner. To improve the synthetic-to-real generalization ability of NeRF, we propose ContraNeRF, a novel approach that gen- eralizes well from synthetic data to real data via contrastive learning with geometry consistency. In many 2D vision tasks, contrastive learning has been shown to improve the generalization ability of models [20, 42, 43] by enhancing the consistency of positive pairs. In 3D scenes, geometry is related to multi-view appearance consistency [24, 37], and contrastive learning may help models predict accurate ge- ometry by enhancing multi-view consistency. In this paper, we propose geometry-aware contrastive learning to learn a multi-view consistent features representation by compar- ing the similarities of local features for each pair of sourceviews (see Fig.3). Specifically, for pixels of each source view, we first aggregate information along the ray projected to other source views to get the geometry-enhanced fea- tures. Then, we sample a batch of target pixels from each source view as the training batch for contrastive learning and project them to other views to get positive and neg- ative samples. The InfoNCE loss [35] is calculated in a weighted manner. Finally, we render the ray by learning a general view interpolation function following [37]. Ex- periments show that when trained on the synthetic data, our method outperforms the recent concurrent generalizable NeRF works [4, 19, 24, 37, 46] and can render high-quality novel view while preserving fine-grained details for unseen scenes. Moreover, under the real-to-real setting, our method also performs better than existing neural radiance field gen- eralization methods. In summary, our contributions are: 1. Investigate the effects of synthetic data in NeRF-based novel view synthesis and observe that models trained on synthetic data tend to predict sharper but less accu- rate volume densities when tested on real data; 2. Propose geometry-aware contrastive learning to learn multi-view consistent features with geometric con- straints, which significantly improves the model’s synthetic-to-real generalization ability; 3. Our method achieves state-of-the-art results for gener- alizable novel view synthesis under both synthetic-to- real and real-to-real settings.
Zhang_Towards_Unsupervised_Object_Detection_From_LiDAR_Point_Clouds_CVPR_2023	Abstract In this paper, we study the problem of unsupervised ob- ject detection from 3D point clouds in self-driving scenes. We present a simple yet effective method that exploits (i) point clustering in near-range areas where the point clouds are dense, (ii) temporal consistency to ﬁlter out noisy unsu- pervised detections, (iii) translation equivariance of CNNs to extend the auto-labels to long range, and (iv) self- supervision for improving on its own. Our approach, OYS- TER (Object Discover yviaSpatio- Temporal Reﬁnement), does not impose constraints on data collection (such as repeated traversals of the same location), is able to de- tect objects in a zero-shot manner without supervised ﬁne- tuning (even in sparse, distant regions), and continues to self-improve given more rounds of iterative self-training. To better measure model performance in self-driving scenar- ios, we propose a new planning-centric perception metric based on distance-to-collision. We demonstrate that our unsupervised object detector signiﬁcantly outperforms un- supervised baselines on PandaSet and Argoverse 2 Sen- sor dataset, showing promise that self-supervision com- bined with object priors can enable object discovery in the wild. For more information, visit the project website: https://waabi.ai/research/oyster .	1. Introduction When a large set of annotations are available, supervised learning can solve the task of 3D object detection remark- ably well thanks to the power of neural networks, as proven by many successful object detectors developed in the past decade [24,36,47,70]. However, since most existing data is unlabeled, human annotations are currently the bottleneck for data-driven learning algorithms, as they require tedious manual effort that is very costly in practice. While there has been some effort on using weaker supervision to train ob- ject detectors [4, 57, 68], it is worth noting that human and animal brains are able to perceive objects without explicit yWork done at Waabi. Mengye is now at New York University. Near-range: dense point cloud with clear clusters Long-range: sparse point cloud Figure 1. Visualization of a sequence of ﬁve 3D point clouds. At near range, the point clouds are very dense and we can clearly distinguish clusters, motivating us to use prior knowledge. At far range, the point clouds are quite sparse and the objects do not ap- pear as obvious, leading us to explore zero-shot generalization. labels at all [2]. This naturally inspires us to ask whether we can design unsupervised learning algorithms that dis- cover objects from raw streams of sensor data on their own. Unsupervised object detection has long been studied in computer vision, albeit in various forms. For instance, nu- merous ways of unsupervised object proposals were consid- ered as the ﬁrst stage of an object detector [20]. These meth- ods leverage a variety of cues including colors and edge boundaries [1], graph structures [18], and motion cues [50]. While those methods are no longer popular in today’s object detectors due to end-to-end supervised training, they offer important intuitions of what an object is . In recent years, unsupervised object detection has made a comeback under the name of object-centric models [6,15, 33, 34, 38, 39, 58]. The essential idea is to train an auto- encoder with a structured decoder such that the network is forced to decompose the scene into a set of individual objects during reconstruction. Most of those models only show experiments on synthetic toy datasets, where the back- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 9317 ground lacks any ﬁne-grained details, and the foreground objects have simplistic shapes and distinct colors. The rea- son why object-centric models have struggled to scale to re- alistic data is that their mechanism of object decomposition is based on careful balancing of model capacities between the foreground and background modules. Consequently, an increase in model capacity, which is needed for real-world data, breaks the brittle balance between modules and result in a failure in scene decomposition. Unsupervised object discovery in the wild remains an open challenge. In this work, we study unsupervised object detection from point clouds in the context of self-driving vehicles (SDVs). This is a challenging task due to occlusion as well as the sparsity of the observations particularly at range. We refer the reader to Fig. 1 for an example. Despite its appeal, unsupervised object detection from LiDAR has received lit- tle attention. Recent work [66] exploits repeated traver- sals of the same region to understand the persistence of a point over time and discover mobile objects from that in- formation. However, the assumption of repeated traversals in acquiring point clouds restricts the applicability of the method. Instead, we study the most generic setting of unsu- pervised detection given any raw sequence of point clouds. Our method OYSTER (Object Discover yviaSpatio- Temporal Reﬁnement ) carefully combines key ideas from density-based spatial clustering, temporal consistency, equivariance and self-supervised learning in a uniﬁed framework that exploits their strengths while overcoming their shortcomings. Firstly, we exploit point clustering to obtain initial pseudo-labels to bootstrap an object detector in the near range, where point density is high (see Fig. 1). We then employ unsupervised tracking to ﬁlter out tempo- rally inconsistent objects. Since point clustering does not work well in the long-range where observations are sparse, we exploit the translation equivariance of CNNs to train on high-quality, near-range pseudo-labels and zero-shot gen- eralize to long range. To bridge the density gap between short and long-range at training vs. inference, we propose a novel random LiDAR ray dropping strategy. Finally, we design a self-improvement loop in which this bootstrapped model can self-train. At every round of self-improvement, we utilize the temporal consistency of objects to automati- cally reﬁne the detections from the model at the previous iteration, and use these reﬁned outputs as pseudo-labels for training. Our experiments on Pandaset [63] and Ar- goverse V2 Sensor [60] demonstrate that OYSTER clearly outperforms other unsupervised methods, both under stan- dard metrics based on intersection-over-union (IoU) and our proposed metric based on distance-to-collision (DTC). We hope that our work serves as a step towards building percep- tion systems that automatically improve with more data and compute without being bottlenecked by human supervision.
Zhang_Generating_Human_Motion_From_Textual_Descriptions_With_Discrete_Representations_CVPR_2023	Abstract In this work, we investigate a simple and must-known con- ditional generative framework based on Vector Quantised- Variational AutoEncoder (VQ-VAE) and Generative Pre- trained Transformer (GPT) for human motion generation from textural descriptions. We show that a simple CNN- based VQ-VAE with commonly used training recipes (EMA and Code Reset) allows us to obtain high-quality discrete rep- resentations. For GPT, we incorporate a simple corruption strategy during the training to alleviate training-testing dis- crepancy. Despite its simplicity, our T2M-GPT shows better performance than competitive approaches, including recent diffusion-based approaches. For example, on HumanML3D, which is currently the largest dataset, we achieve compara- ble performance on the consistency between text and gener- ated motion (R-Precision), but with FID 0.116 largely outper- forming MotionDiffuse of 0.630. Additionally, we conduct analyses on HumanML3D and observe that the dataset size is a limitation of our approach. Our work suggests that VQ- VAE still remains a competitive approach for human motion generation. Our implementation is available on the project page: https://mael-zys.github.io/T2M-GPT/ .	1. Introduction Generating motion from textual descriptions can be used in numerous applications in the game industry, film-making, and animating robots. For example, a typical way to access new motion in the game industry is to perform motion cap- ture, which is expensive. Therefore automatically generating motion from textual descriptions, which allows producing meaningful motion data, could save time and be more eco- nomical. Motion generation conditioned on natural language is challenging, as motion and text are from different modali- ties. The model is expected to learn precise mapping from Figure 1. Visual results on HumanML3D [22]. Our approach is able to generate precise and high-quality human motion consistent with challenging text descriptions. More visual results are on the project page. the language space to the motion space. To this end, many works propose to learn a joint embedding for language and motion using auto-encoders [3, 21, 65] and V AEs [52, 53]. MotionClip [65] aligns the motion space to CLIP [55] space. ACTOR [52] and TEMOES [53] propose transformer-based V AEs for action-to-motion and text-to-motion respectively. These works show promising performances with simple de- scriptions and are limited to producing high-quality motion when textual descriptions become long and complicated. Guo et al. [22] and TM2T [23] aim to generate motion This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 14730 sequences with more challenging textual descriptions. How- ever, both approaches are not straightforward, involve three stages for text-to-motion generation, and sometimes fail to generate high-quality motion consistent with the text (See Figure 4 and more visual results on the project page). Re- cently, diffusion-based models [32] have shown impressive results on image generation [60], which are then introduced to motion generation by MDM [66] and MotionDiffuse [74] and dominates text-to-motion generation task. However, we find that compared to classic approaches, such as VQ- V AE [68], the performance gain of the diffusion-based ap- proaches [66, 74] might not be that significant. In this work, we are inspired by recent advances from learning the discrete representation for generation [5,15,16,19,45,58,68,70] and investigate a simple and classic framework based on Vector Quantized Variational Autoencoders (VQ-V AE) [68] and Generative Pre-trained Transformer (GPT) [56, 69] for text- to-motion generation. Precisely, we propose a two-stage method for motion generation from textual descriptions. In stage 1, we use a standard 1D convolutional network to map motion sequences to discrete code indices. In stage 2, a standard GPT-like model [56, 69] is learned to generate sequences of code in- dices from pre-trained text embedding. We find that the naive training of VQ-V AE [68] suffers from code collapse. One effective solution is to leverage two standard recipes during the training: EMA andCode Reset . We provide a full anal- ysis of different quantization strategies. For GPT, the next token prediction brings inconsistency between the training and inference. We observe that simply corrupting sequences during the training alleviates this discrepancy. Moreover, throughout the evolution of image generation, the size of the dataset has played an important role. We further explore the impact of dataset size on the performance of our model. The empirical analysis suggests that the performance of our model can potentially be improved with larger datasets. Despite its simplicity, our approach can generate high- quality motion sequences that are consistent with challeng- ing text descriptions (Figure 1 and more on the project page). Empirically, we achieve comparable or even better performances than concurrent diffusion-based approaches MDM [66] and HumanDiffuse [74] on two widely used datasets: HumanML3D [22] and KIT-ML [54]. For example, on HumanML3D, which is currently the largest dataset, we achieve comparable performance on the consistency between text and generated motion (R-Precision), but with FID 0.116 largely outperforming MotionDiffuse of 0.630. We conduct comprehensive experiments to explore this area, and hope that these experiments and conclusions will contribute to future developments. In summary, our contributions include: •We present a simple yet effective approach for mo- tion generation from textual descriptions. Our ap-proach achieves state-of-the-art performance on Hu- manML3D [22] and KIT-ML [54] datasets. •We show that GPT-like models incorporating discrete representations still remain a very competitive approach for motion generation. •We provide a detailed analysis of the impact of quanti- zation strategies and dataset size. We show that a larger dataset might still offer a promising prospect to the community. Our implementation is available on the project page.
Yu_X-Pruner_eXplainable_Pruning_for_Vision_Transformers_CVPR_2023	Abstract Recently vision transformer models have become promi- nent models for a range of tasks. These models, how- ever, usually suffer from intensive computational costs and heavy memory requirements, making them impractical for deployment on edge platforms. Recent studies have pro- posed to prune transformers in an unexplainable manner, which overlook the relationship between internal units of the model and the target class, thereby leading to infe- rior performance. To alleviate this problem, we propose a novel explainable pruning framework dubbed X-Pruner, which is designed by considering the explainability of the pruning criterion. Speciﬁcally, to measure each prunable unit’s contribution to predicting each target class, a novel explainability-aware mask is proposed and learned in an end-to-end manner. Then, to preserve the most informative units and learn the layer-wise pruning rate, we adaptively search the layer-wise threshold that differentiates between unpruned and pruned units based on their explainability- aware mask values. To verify and evaluate our method, we apply the X-Pruner on representative transformer mod- els including the DeiT and Swin Transformer. Comprehen- sive simulation results demonstrate that the proposed X- Pruner outperforms the state-of-the-art black-box methods with signiﬁcantly reduced computational costs and slight performance degradation. Code is available at https: //github.com/vickyyu90/XPruner .	1. Introduction Over the last few years, transformers have attracted in- creasing attention in various challenging domains, such as natural language processing, vision, or graphs [3, 9]. It is composed of two key modules, namely the Multi-Head At- tention (MHA) and Multi-Layer Perceptron (MLP). How- ever, similar to CNNs, the major limitations of transform- ers include the gigantic model sizes with intensive compu- tational costs. Which severely restricts their deployment *Corresponding author.in resource-constrained devices like edge platforms. To compress and accelerate transformer models, a variety of techniques naturally emerge. Popular approaches include weight quantization [30], knowledge distillation [28], ﬁlter compression [24], and model pruning [21]. Among them, model pruning especially structured pruning has gained considerable interest that removes the least important pa- rameters in pre-trained models in a hardware-friendly man- ner, which is thus the focus of our paper. Due to the signiﬁcant structural differences between CNNs and transformers, although there is prevailing suc- cess in CNN pruning methods, the research on pruning transformers is still in the early stage. Existing studies could empirically be classiﬁed into three categories. (1) Criterion-based pruning resorts to preserving the most im- portant weights/attentions by employing pre-deﬁned crite- ria, e.g., the L1/L2 norm [20], or activation values [6]. (2) Training-based pruning retrains models with hand-crafted sparse regularizations [31] or resource constraints [28, 29]. (3) Architecture-search pruning methods directly search for an optimal sub-architecture based on pre-deﬁned policies [3, 10]. Although these studies have made considerable progress, two fundamental issues have not been fully ad- dressed, i.e., the optimal layer-wise pruning ratio and the weight importance measurement. For the ﬁrst issue, the ﬁnal performance is notably af- fected by the selection of pruning rates for different layers. To this end, some relevant works have proposed a series of methods for determining the optimal per-layer rate [7, 11]. For instance, Michel et al. [18] investigate the effectiveness of attention heads in transformers for NLP tasks and pro- pose to prune attention heads with a greedy algorithm. Yu et al. [28] develop a pruning algorithm that removes attention scores below a learned per-layer threshold while preserving the overall structure of the attention mechanism. However, the proposed methods do not take into account the inter- dependencies between weight. Recently, Zhu et al. [31] in- troduce the method VTP with a sparsity regularization to identify and remove unimportant patches and heads from the vision transformers. However, VTP needs to try the thresholds manually for all layers. 1 This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 24355 For the second issue, previous studies resort to identi- fying unimportant weights by various importance metrics, including magnitude-based, gradient-based [12, 19], and mask-based [27]. Among them, the magnitude-based ap- proaches usually lead to suboptimal results as it does not take into account the potential correlation between weights [24]. In addition, gradient-based methods often tend to prune weights with small values, as they have small gra- dients and may not be identiﬁed as important by the back- ward propagation. Finally, the limitation of current mask- based pruning lies in two folds: (1) Most mask-based prun- ing techniques manually assign a binary mask w.r.t. a unit according to a per-layer pruning ratio, which is inefﬁcient and sub-optimal. (2) Most works use a non-differentiable mask, which results in an unstable training process and poor convergence. In this paper, we propose a novel explainable structured pruning framework for vision transformer models, termed X-Pruner, by considering the explainability of the prun- ing criterion to solve the above two problems. As stated in the eXplainable AI (XAI) ﬁeld [2], important weights in a model typically capture semantic class-speciﬁc infor- mation. Inspired by this theory, we propose to effectively quantitate the importance of each weight in a class-wise manner. Firstly, we design an explainability-aware mask for each prunable unit (e.g., an attention head or matrix in linear layers), which measures the unit’s contribution to pre- dicting every class and is fully differentiable. Secondly, we use each input’s ground-truth label as prior knowledge to guide the mask learning, thus the class-level information w.r.t. each input will be fully utilized. Our intuition is that if one unit generates feature representations that make a pos- itive contribution to a target class, its mask value w.r.t. this class would be positively activated, and deactivated other- wise. Thirdly, we propose a differentiable pruning oper- ation along with a threshold regularizer. This enables the search of thresholds through gradient-based optimization, and is superior to most previous studies that prune units with hand-crafted criteria. Meanwhile, the proposed prun- ing process can be done automatically, i.e., discriminative units that are above the learned threshold are retained. In this way, we implement our layer-wise pruning algorithm in an explainable manner automatically and efﬁciently. In summary, the major contributions of this paper are: • We propose a novel explainable structured pruning framework dubbed X-Pruner, which prunes units that make less contributions to identifying all the classes in terms of explainability. To the best knowledge of the authors, this is the ﬁrst work to develop an explainable pruning framework for vision transformers; • We propose to assign each prunable unit an explainability-aware mask, with the goal of quantify-ing its contribution to predicting each class. Speciﬁ- cally, the proposed mask is fully differentiable and can be learned in an end-to-end manner; • Based on the obtained explainability-aware masks, we propose to learn the layer-wise pruning thresholds that differentiate the important and less-important units via a differentiable pruning operation. Therefore, this pro- cess is done in an explainable manner; • Comprehensive simulation results are presented to demonstrate that the proposed X-Pruner outperforms a number of state-of-the-art approaches, and shows its superiority in gaining the explainability for the pruned model.
Xu_Video_Dehazing_via_a_Multi-Range_Temporal_Alignment_Network_With_Physical_CVPR_2023	Abstract Video dehazing aims to recover haze-free frames with high visibility and contrast. This paper presents a novel framework to effectively explore the physical haze priors and aggregate temporal information. Specifically, we de- sign a memory-based physical prior guidance module to encode the prior-related features into long-range memory. Besides, we formulate a multi-range scene radiance recov- ery module to capture space-time dependencies in multiple space-time ranges, which helps to effectively aggregate tem- poral information from adjacent frames. Moreover, we con- struct the first large-scale outdoor video dehazing bench- mark dataset, which contains videos in various real-world scenarios. Experimental results on both synthetic and real conditions show the superiority of our proposed method.	1. Introduction Haze largely degrades the visibility and contrast of the outdoor scenes, which adversely affects the performance of downstream vision tasks, such as the detection and segmen- tation in autonomous driving and surveillance. According to the atmospheric scattering model [18, 38], the formation of a hazy image is described as: I(x) =J(x)t(x) +A(1−t(x)), (1) where I, J, A, t denote the observed hazy image, scene radi- ance, atmospheric light, and transmission, respectively, and xis the pixel index. The transmission t=e−βd(x)describes the scene radiance attenuation caused by the light scattering, where βis the scattering coefficient of the atmosphere, and ddenotes the scene depth. Video dehazing benefits from temporal clues, such as highly correlated haze thickness and lighting conditions, as ⋆: This work was done during Jiaqi Xu’s internship at Shanghai Artificial Intelligence Laboratory. : Corresponding authors ([email protected]; [email protected]). (a) Hazy frame (b) VDH [46] (c) CG-IDN [59] (d) Our method Figure 1. Visual comparison on a real-world hazy video. Our method trained on our outdoor dehazing dataset clearly removes haze without color distortion. well as the moving foreground objects and backgrounds. Early deep learning-based video dehazing methods lever- age temporal information by simply concatenating input frames or feature maps [27, 46]. Recently, GC-IDN [59] proposes to use cost volume and confidence to align and ag- gregate temporal information. However, existing video de- hazing methods suffer from several limitations. First, these approaches either obtain haze-free frames from the phys- ical model-based component estimation [27, 46] or ignore the explicit physical prior embedded in the haze imaging model [59]. The former suffers from inaccurate interme- diate prediction, thus leading to error accumulation in the final results, while the latter overlooks the physical prior in- formation, which plays an important role in haze estimation and scene recovery. Second, these methods aggregate tem- poral information by using input/feature stacking or frame- to-frame alignment in a local sliding window, which is hard to obtain global and long-range temporal information. In this work, we present a novel video dehazing frame- work via a Multi-range temporal Alignment network with Physical prior (MAP-Net) to address the aforementioned is- sues. First, we design a memory-based physical prior guid- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 18053 ance module, which aims to inject the physical prior to help the scene radiance recovery. Specifically, we perform fea- ture disentanglement according to the physical model with two decoders, where one estimates the transmission and atmospheric light, and the other recovers scene radiance. The feature extracted from the first decoder is leveraged as the physical haze prior, which is integrated into the second decoder for scene radiance recovery. To infer the global physical prior in a long-range video, we design a physical prior token memory that effectively encodes prior-related features into compact tokens for efficient memory reading. Second, we introduce a multi-range scene radiance re- covery module to capture space-time dependencies in mul- tiple space-time ranges. This module first splits the adjacent frames into multiple ranges, then aligns and aggregates the corresponding recurrent range features, and finally recov- ers the scene radiance. Unlike CG-IDN [59], which aligns the adjacent features frame-by-frame, we align the features of adjacent frames into multiple sets with different ranges, which helps to explore the temporal haze clues in various time intervals. We further design a space-time deformable attention to warp the features of multiple ranges to the target frame, followed by a guided multi-range complementary in- formation aggregation. Also, we use an unsupervised flow loss to encourage the network to focus on the aligned areas and train the whole network in an end-to-end manner. In addition, the existing learning-based video dehaz- ing methods are mainly trained and evaluated on indoor datasets [27, 46, 59], which suffer from performance degra- dation in real-world outdoor scenarios. Thus, we construct an outdoor video dehazing benchmark dataset, HazeWorld, which has three main properties. First, it is a large-scale synthetic dataset with 3,588 training videos and 1,496 test- ing videos. Second, we collect videos from diverse out- door scenarios, e.g., autonomous driving and life scenes. Third, the dataset has various downstream tasks for eval- uation, such as segmentation and detection. Various ex- periments on both synthetic and real datasets demonstrate the effectiveness of our approach, which clearly outper- forms the existing image and video dehazing methods; see Fig. 1. The code and dataset are publicly available at https://github.com/jiaqixuac/MAP-Net . Our main contributions are summarized as follows: • We present a novel framework, MAP-Net, for video dehazing. A memory-based physical prior guidance module is designed to enhance the scene radiance re- covery, which encodes haze-prior-related features into long-range memory. • We introduce a recurrent multi-range scene radiance recovery module with the space-time deformable at- tention and the guided multi-range aggregation, which effectively captures long-range temporal haze and scene clues from the adjacent frames.• We construct a large-scale outdoor video dehazing dataset with diverse real-world scenarios and labels for downstream task evaluation. • Extensive experiments on both synthetic and real con- ditions demonstrate our superior performance against the recent state-of-the-art methods.
Zeng_Distilling_Focal_Knowledge_From_Imperfect_Expert_for_3D_Object_Detection_CVPR_2023	Abstract Multi-camera 3D object detection blossoms in recent years and most of state-of-the-art methods are built up onthe bird’s-eye-view (BEV) representations. Albeit remark-able performance, these works suffer from low efﬁciency.Typically, knowledge distillation can be used for modelcompression. However , due to unclear 3D geometry reason-ing, expert features usually contain some noisy and confus-ing areas. In this work, we investigate on how to distill theknowledge from an imperfect expert. We propose FD3D, aF ocal Distiller for 3D object detection. Speciﬁcally, a set ofqueries are leveraged to locate the instance-level areas formasked feature generation, to intensify feature representa-tion ability in these areas. Moreover , these queries searchout the representative ﬁne-grained positions for reﬁned dis-tillation. We verify the effectiveness of our method by ap-plying it to two popular detection models, BEVF ormer andDETR3D. The results demonstrate that our method achievesimprovements of 4.07 and 3.17 points respectively in termsof NDS metric on nuScenes benchmark. Code is hostedathttp s:// github.com/OpenPerceptionX/ BEVPerception-Survey-Recipe .	1. Introduction Accurate 3D object detection is a vital component in au- tonomous driving. To achieve this, most methods [ 14,37] resort to LiDAR sensors and dominate the public bench-marks [ 1,27]. Despite the performance gap, pure vision approaches are still worthy of in-depth inquiry, since cam-eras can provide rich semantic information and are low-costand easy-to-deploy. Among these, bird’s-eye-view (BEV) detection has drawn extensive attention from both industryand academia, and shown great potential to narrow downthe performance gap [ 15,21]. However, such models tend to be computationally consuming. *Corresponding author at [email protected] (a) Global distillation (b) Attentive distillation (c) Generative distillation with random mask(d) Generative and focal distillation (ours)Random maskForeground region Generator Query mask GeneratorExpert Abstraction Apprentice Representation Query offset Figure 1. Illustration of the proposed generative and focal distilla- tion method. Compared with others, the proposed manner in 1(d) leverages queries to generate instance masks for masked genera- tive distillation, rather than random masks in 1(c). Moreover, the queries meanwhile search for the representative position to per-form reﬁned distillation, where the distillation region selection ismore ﬁne-grained and ﬂexible than 1(b). In common practice, knowledge distillation can com- press the model and is usually applied to alleviate com-putation overhead. One possible solution is to utilize theLiDAR-based model as the expert [ 4,17], but this requires complex spatial-temporal calibrations and also needs tohandle heterogeneous problems from different modalities.An intuitive question is, can we distill these models solely based on camera sensors? In this work, we intend to ad-dress this problem and focus on the camera-only distillation setting. To the best of our knowledge, our work is the ﬁrstsolution tailored for this setting. Distillation methods in 2D object detection have derived This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 992 Figure 2. Visualization of the predicted bounding boxes and bird’s eye view feature of BEVFormer [ 19]. At the center is the au- tonomous vehicle. In the left subﬁgure, green and blue boxes de- note the ground truth and predictions respectively. In the right sub-ﬁgure, the BEV feature is ray-shaped and contains a lot of noise.Areas with incorrect high activation appear behind objects due toocclusion, which easily introduces false positives, e.g., the region circled by the ellipse. various types as depicted in Fig. 1(a)-(c), but their effec- tiveness is not veriﬁed in 3D object detection. The mainchallenge in camera-to-camera distillation for 3D object de- tection comes from the imperfect expert features. Due tothe lack of accurate 3D information, expert features drawn from 2D images usually contain some noisy and confusingareas. To better illustrate this point, we visualize the BEVfeatures from BEVFormer [ 19] in Fig. 2. The unclear oc- clusion reasoning makes the BEV features suffer from theray-shape artifacts. The imperfect BEV features result inmany false positives. Directly mimicking these featuresfrom experts such as Fig. 1(a) may exacerbate the draw- back. Some 2D detection distillation methods [ 29] propose to focus on foreground-orient regions, as shown in Fig. 1 (b). However, background regions are also important as proved in [ 7,35,13]. We categorize these methods as at- tentive distillation. Compared with 2D object detection, theimbalance between foreground and background in 3D ar-eas is much more severe. Balancing these weights is not ageneral solution. The latest study MGD [ 36] demonstrates the effectiveness of masked image modeling (MIM) distil-lation as depicted in Fig. 1(c). However, the global random mask cannot greatly enhance 3D object detection, which isvalidated in Tab. 4a. To this end, we propose a Focal Distiller for 3D ob- ject detection, shortened as FD3D . The schematic diagram of FD3D is shown in Fig. 1(d). Speciﬁcally, A set of queries are leveraged to locate the instance-level focal re-gions, masked generative distillation is performed withinthe regions. Moreover, these queries dynamically searchﬁne-grained representative positions for focal distillation.Two complementary modules guide the apprentice networkto generate enhanced feature representation on focal re-gions. In summary, our work makes three-fold contribu-tions:1. To the best of our knowledge, this is the ﬁrst work to explore knowledge distillation on the camera-only 3Dobject detection. We reveal the challenge relies on howto distill focal knowledge from an imperfect 3D objectdetector expert. 2. We propose FD3D, which utilizes a set of queries to distill focal knowledge. With these queries, coarse-grained focal regions are selected for masked gener-ation, and ﬁne-grained focal regions are searched outfor instance-oriented reﬁnement distillation. 3. FD3D serves as a plug-and-play module. It can be eas- ily extended to various detectors. The improvementswith 4.07 and 3.17 NDS can be obtained with FD3Dassembled in BEVFormer and DETR3D, respectively.
Zhang_Learning_To_Generate_Language-Supervised_and_Open-Vocabulary_Scene_Graph_Using_Pre-Trained_CVPR_2023	Abstract Scene graph generation (SGG) aims to abstract an im- age into a graph structure, by representing objects as graph nodes and their relations as labeled edges. However, two knotty obstacles limit the practicability of current SGG methods in real-world scenarios: 1) training SGG mod- els requires time-consuming ground-truth annotations, and 2) the closed-set object categories make the SGG mod- els limited in their ability to recognize novel objects out- side of training corpora. To address these issues, we nov- elly exploit a powerful pre-trained visual-semantic space (VSS) to trigger language-supervised and open-vocabulary SGG in a simple yet effective manner. Specifically, cheap scene graph supervision data can be easily obtained by parsing image language descriptions into semantic graphs. Next, the noun phrases on such semantic graphs are di- rectly grounded over image regions through region-word alignment in the pre-trained VSS. In this way, we enable open-vocabulary object detection by performing object cat- egory name grounding with a text prompt in this VSS. On the basis of visually-grounded objects, the relation repre- sentations are naturally built for relation recognition, pur- suing open-vocabulary SGG. We validate our proposed ap- proach with extensive experiments on the Visual Genome benchmark across various SGG scenarios (i.e., supervised / language-supervised, closed-set / open-vocabulary). Con- sistent superior performances are achieved compared with existing methods, demonstrating the potential of exploiting pre-trained VSS for SGG in more practical scenarios.	1. Introduction Scene graph [10] is a structured representation for de- scribing image semantics. It abstracts visual objects as graph nodes and represents their relations as labeled graph edges. The task of scene graph generation ( SGG ) [6, 14, *Corresponding author Phrase grounding A man holds a baseball bat on the ground .Semantic parsingman baseball batgroundon hold player baseball batgroundon hold Word embedding Region embedding Novel classVisual -Semantic SpaceVisually -grounded semantic graphImage language description man Open -vocabulary generalizationplayerImagea baseball bat. ground. ... player . Text prompt inputPredicted scene graph bUsed as weak scene graph supervision Novel objectFigure 1. An illustration of exploiting a pre-trained visual- semantic space (VSS) to trigger language-supervised and open- vocabulary scene graph generation (SGG). (a) We acquire weak scene graph supervision by semantically parsing the image lan- guage description and grounding noun phrases on image re- gions via VSS. (b) At SGG inference time, thanks to the open- vocabulary generalization naturally rooted in VSS, the novel ob- ject name (e.g., player) in the text prompt input can be well aligned to one image region, which is regarded as its detection. 20, 26, 40, 47, 48, 50, 51, 57, 60, 63, 64] plays an important role for fine-grained visual understanding, which has shown promising results in facilitating various downstream appli- cations, such as image-text retrieval [24,38,49], image cap- tioning [2,22,32,35,52,54,55,66], cross-media knowledge graph construction [18, 45] and robot planning [1]. Though great effort has been made, SGG of the current stage still faces two knotty obstacles that limit its practi- cability in real-world scenarios. 1) Training SGG mod- els requires massive ground-truth scene graphs that are ex- pensive for manual annotation. Annotators have to draw bounding boxes for all objects in an image and connect possible interacted object pairs, and assign object/relation labels. Since assigned labels might be ambiguous, further verification and canonicalization processing are usually re- quired [14]. Finally, a scene graph in the form of a set of ⟨subject, predicate, object ⟩triplets with subject and ob- ject bounding boxes is constructed. Such annotating pro- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 2915 cess is time-consuming and tedious, costing much human labor and patience. 2) Almost all existing SGG methods [20,21,26,47,48,50,51,60] involve a pre-defined closed set of object categories, making them limited in recognizing novel objects outside of training corpora. However, real- world scenes contain a boarder set of visual concepts than any pre-defined category pool. It is very likely to encounter unseen/novel categories. When this happens, current SGG models either classify novel objects to a known category or fail to detect them like background regions. Accordingly, the prediction of their interactions/relations with other ob- jects is negatively affected or just neglected. This may lead to problems. For example, a real-world robot may take inap- propriate actions using such closed-set SGG models [1,42]. Recently, there is a trend of leveraging free-form lan- guage supervision for benefiting visual recognition tasks via large-scale language-image pre-training [7, 15, 17, 36, 53, 59, 67]. These methods (e.g., CLIP [36]) perform pre-training on massive easily-obtained image-text pairs to learn a visual-semantic space (VSS), and have demon- strated great zero-shot transferability. Especially, the re- cent grounded language-image pre-training (GLIP) [17] has learned an object-level and semantic-rich VSS. Based on the learned VSS, it has established new state-of-the-art per- formances in phrase grounding and zero-shot object de- tection. This indicates such pre-trained VSS has power- ful multi-modal alignment ability (i.e., image regions and text phrases that have similar semantics get close embed- dings) and open-vocabulary generalization ability (i.e., cov- ering virtually any concepts in the pre-training image-text corpus). This inspires our thought of addressing the afore- mentioned obstacles in SGG using the pre-trained VSS. On the one hand, taking advantage of its multi-modal align- ment ability, we can cheaply acquire scene graph supervi- sion from an image description (e.g., retrieving image re- gions aligned with noun phrases and re-arranging the de- scription into a scene-graph-like form). On the other hand, by leveraging its open-vocabulary generalization ability, it is promising to enable novel category prediction in SGG. In this work, we investigate the opportunity of fully ex- ploiting the VSS learned by language-image pre-training to trigger language-supervised and open-vocabulary SGG. Specifically, we obtain weak scene graph supervision by se- mantically parsing an image language description into a se- mantic graph, then grounding its noun phrases over image regions through region-word alignment in the pre-trained VSS (Figure 1 (a)). Moreover, we propose a novel SGG model, namely Visual-Semantic Space for Scene graph gen- eration ( VS3). It takes a raw image and a text prompt con- taining object category names as inputs, and projects them into the shared VSS as embeddings. Next, VS3performs object detection by aligning the embeddings of category names and image regions. Based on high-confidence de-tected objects, VS3builds relation representations for object pairs with a devised relation embedding module that fully mines relation patterns from visual and spatial perspec- tives. Finally, a relation prediction module takes relation representations to infer relation labels. The predicted scene graph is composed by combining object detections and in- ferred relation labels. During training, visually-grounded semantic graphs parsed from image descriptions could be used as weak scene graph supervision, achieving language- supervised SGG. At SGG inference time, when using a text prompt input containing novel categories, VS3manages to detect novel objects thanks to the open-vocabulary general- ization ability natually rooted in VSS, hence allowing for open-vocabulary SGG (Figure 1 (b)). In summary, we have made the following contributions: (1) the exploitation of a pre-trained VSS provides an el- egant solution for addressing obstacles to triggering both language-supervised and open-vocabulary SGG, making a solid step toward real-world usage of SGG. (2) The pro- posed VS3model is a new and versatile framework, which effectively transfers language-image pre-training knowl- edge for benefiting SGG. (3) We fully validate the effective- ness of our approach through extensive experiments on the Visual Genome benchmark, and have set new state-of-the- art performances spanning across all settings (i.e., super- vised / language-supervised, closed-set / open-vocabulary).
Yu_Overcoming_the_Trade-Off_Between_Accuracy_and_Plausibility_in_3D_Hand_CVPR_2023	Abstract Direct mesh ﬁtting for 3D hand shape reconstruction is highly accurate. However , the reconstructed meshes areprone to artifacts and do not appear as plausible hand shapes. Conversely, parametric models like MANO ensureplausible hand shapes but are not as accurate as the non- parametric methods. In this work, we introduce a novelweakly-supervised hand shape estimation framework that integrates non-parametric mesh ﬁtting with MANO modelin an end-to-end fashion. Our joint model overcomes the tradeoff in accuracy and plausibility to yield well-alignedand high-quality 3D meshes, especially in challenging two- hand and hand-object interaction scenarios.	1. Introduction State-of-the-art monocular RGB-based 3D hand recon- struction methods [ 6,10,17,21,22,28] focus on recover- ing highly accurate 3D hand meshes. As accuracy is mea- sured by an average joint or vertex position error, recov-ered hand meshes may be well-aligned in 3D space but still be physically implausible. The 3D mesh surface may have irregular protrusions or collapsed regions (see Fig. 1), especially around the ﬁngers. The meshes may also suf- fer from incorrect contacts or penetrations when there arehand-object or two-hand interactions. Yet methods that pri- oritize physical plausibility, especially in interaction set-tings [ 3,8,10,14,20,21], are signiﬁcantly less accurate in 3D alignment. In summary, the current body of work predom- inantly favours either 3D alignment accuracy or physicalplausibility, but cannot achieve both. A closer examination reveals that the trade-off between 3D alignment and plausibility is also split methodology- wise. Methods that use the MANO model [ 30] produce plausible hand poses and hand shapes [ 2,7,38,40,42] due to the statistical parameterization of MANO. However, it is challenging to directly regress these parameters, since Figure 1. The vertex error vs. edge length error indicates that existing methods trade-off alignment accuracy with plausibility. MANO-based methods (triangles) vs. non-parametric model- based methods (circles) have a trade-off between vertex error andedge length error; our combined method (stars) can overcome thistrade-off to yield well-aligned and plausible meshes. Plot of re-sults from InterHand2.6M [ 27]; visualization from FreiHand [ 42]. the mapping from an image to the MANO parameter space is highly non-linear. As a consequence, MANO-basedmethods lag in 3D alignment accuracy compared to non-parametric methods. Non-parametric methods [ 6,10,11,17,18,21,22,28] di- rectly ﬁt a 3D mesh to image observations. Direct meshﬁtting is accurate but is prone to surface artifacts. In scenar- ios with hand-object or hand-hand interactions, mesh pen- etrations cannot be resolved meaningfully even with regu-larizers such as contact losses [ 14] due to the unconstrained This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 544 optimization. Attention mechanism [ 21,32] can mitigate some penetrations and artifacts, but the inherent problem remains. As such, the favoured approaches for hand-objectand hand-hand interactions are still driven by MANO mod-els [ 3,8,13,14,38]. In this work, we aim to recover high-quality hand meshes that are accurately aligned and have minimal artifacts and penetrations. To avoid a trade-off, we leverage direct meshﬁtting for alignment accuracy and guidance from MANOfor plausibility. Combine the non-parametric and paramet-ric models is straightforward in terms of motivation. How-ever, merging the two is non-trivial because it requires a mapping from non-parametric mesh vertices to parametric model parameters. This mapping, analogous to the map-ping from an RGB image, is highly non-linear and difﬁcult to achieve directly [ 16]. One of our key contributions in this work is a method to accurately map non-parametric hand mesh vertices toMANO joint parameters θ. To do so, we perform a two-step mapping, from mesh-vertices to the joint coordinates, andjoints coordinates to θ. In the literature, the common prac- tice for the former is to leverage the Jmatrix in MANO and linearly regress the joints from the mesh [ 17,20,21]. Yet the Jmatrix was designed to only map MANO-derived meshes to joints in a rest pose (see Eq. 10 in [ 25]). As we show in our experiments, applying Jto non-rest poses introduces a gap of around 2 mm. Furthermore, we postulate that there isa domain gap between the estimated non-parametric meshes and MANO-derived meshes, even if both meshes have the same topology. To close this gap – we propose a V AE cor-rection module, to be applied after the linear regression withJ. To map the recovered joints from the mesh to θ, we use a twist-swing decomposition and analytically compute theθ. It has been shown previously in [ 20] that decomposing joint rotations into twist-swing rotations [ 1] can simplify the estimation of human SMPL [ 25] model pose parame- ters. Inspired by [ 20], we also leverage the decomposition and further verify that the twist angle has minimal impacton the hand. Note that obtaining ground truth labels for hand mesh vertices is non-trivial. Our framework lends itself well for weak-supervision. Since the estimated 3D mesh from thenon-parametric decoder is regressed into 3D joints, it canalso be supervised with 3D joints as weak labels (see Fig2). At the same time, the parametric mesh estimated from these joints can be used as a pseudo-label for learning thenon-parametric mesh vertices. Such a procedure distills the knowledge from the parametric model and is effective with-out ground truth mesh annotations. We name our method WSIM 3D Hand , in reference to Weakly-supervised Self- distillation Integration Model for 3D hand shape recon- struction. Our contributions include: • A novel framework that integrates a parametric and anon-parametric mesh model for accurate and plausible 3D hand reconstruction. • A V AE correction module that closes the overlooked gap between non-parametric and MANO 3D poses; • A weakly-supervised pipeline, competitive to a fully- supervised counterpart, using only 3D joint labels to learn 3D meshes. • Signiﬁcant improvements over state-of-the-art on hand-object or two-hand interaction benchmark datasets, especially in hand-object interaction on DexYCB.
Zhang_Exploring_Intra-Class_Variation_Factors_With_Learnable_Cluster_Prompts_for_Semi-Supervised_CVPR_2023	Abstract Semi-supervised class-conditional image synthesis is typically performed by inferring and injecting class labelsinto a conditional Generative Adversarial Network (GAN).The supervision in the form of class identity may be inade-quate to model classes with diverse visual appearances. Inthis paper , we propose a Learnable Cluster Prompt-based GAN (LCP-GAN) to capture class-wise characteristics andintra-class variation factors with a broader source of su- pervision. To exploit partially labeled data, we perform softpartitioning on each class, and explore the possibility of as- sociating intra-class clusters with learnable visual concepts in the feature space of a pre-trained language-vision model, e.g., CLIP . F or class-conditional image generation, we de-sign a cluster-conditional generator by injecting a combi-nation of intra-class cluster label embeddings, and further incorporate a real-fake classiﬁcation head on top of CLIPto distinguish real instances from the synthesized ones, con- ditioned on the learnable cluster prompts. This signiﬁcant-ly strengthens the generator with more semantic language supervision. LCP-GAN not only possesses superior gen- eration capability but also matches the performance of the fully supervised version of the base models: BigGAN andStyleGAN2-ADA, on multiple standard benchmarks.	1. Introduction Generative Adversarial Networks (GANs) have achieved considerable success in modeling complex data distribu- tions and generating high-ﬁdelity images from random vec- tors [ 2,18,25]. To control class semantics in the generation ∗Joint ﬁrst authors. †Corresponding author. Learnable prompt: “ ? ? ? ? [class_name]” “? ? ? ?”ൎ“Wavy Hair” “? ? ? ?” ൎ“Smiling” Class Intra-class cluster Figure 1. Different from generic class-conditional GANs con- ditioned on discrete class labels, LCP-GAN learns intra-classcluster-speciﬁc prompts to guide the generation process and cap-ture underlying variation factors. process, object category is typically represented in the form of a discrete label, which is injected into both generator anddiscriminator through learnable embedding layers. Howev-er, sufﬁcient labeled training data may be difﬁcult to collectin real-world applications. Signiﬁcant efforts have been de-voted to semi-supervised generative learning that aims to re-duce the dependence of class-conditional GANs on labeled training data [ 6,15,29,33]. In the semi-supervised setting, the amount of unlabeled training samples can be signiﬁcantly greater than that of la-beled ones. As one of the early attempts, CatGAN [ 44] trained a discriminator to infer the class labels of real im- ages with high conﬁdence, but not for the synthesized ones. Both TripleGAN [ 29] andΔ-GAN [ 15] incorpo- rated an auxiliary classiﬁer to focus on class label pre- diction in the adversarial training process. The unlabeled images with pseudo labels were used to train the class- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 7392 ٖ ْ ْ ࢍ࢓࢏ࢎ ǥࢉ࢔࢛࢝ ٖ ࢚࢚࢞ࢎ ݖ ࡳ࢘࢕࢚ࢇ࢘ࢋ࢔ࢋࡳ ǥ ݃݊݅ݎݎ݁ܪ ݄ܿ݊݅ܨ݈݈ݑܩ ࢙࢙ࢇ࢒࡯݈݁݌ݎݑܲ ࢋ࢓ࢇࡺ࢒ࢇࢋࡾ ࢙ࢋࢍࢇ࢓ࡵࢋ࢑ࢇࡲ ࢙ࢋࢍࢇ࢓ࡵ ǥǥ ࢋ࢒࢈ࢇ࢔࢘ࢇࢋ࢒ ࢙࢙ࢇ࢒ࢉࡰ࢘࢕࢚ࢇ࢔࢏࢓࢏࢘ࢉ࢙࢏ࡰ࢚࢖࢓࢕࢘࢖ ࢍ࢓࢏ࡱ ࢚࢚࢞ࡱ ߝሺ̴ࣳܿଵሻ ǥ ݑ௭ଵݑ௭ଶǥ݈ٔٔ݅ݑݖܽܮ ݃݊݅ݐ݊ݑܤࣳ ݑ௭௞݈ܴࣳܽ݁ࢍ࢔࢏ࢊࢊࢋ࢈࢓ࡱ Ȁ݁݇ܽܨ ݈ܴܽ݁ Ȁ݁݇ܽܨ ߝሺ̴ࣳܿଶ澼ߝሺܿ௬̴௞ࣳሻ ࣳݒ̴ଵଵࣳݒ̴ଵଶڮࣳݒభோݏݏ̴݈ܽܿ݁݉ܽ݊ Ǥ ࣳݒ̴௞ࣳଵࣳݒ̴௞ࣳଶڮࣳݒೖࣳோݏݏ̴݈ܽܿ݁݉ܽ݊ Ǥ෍ ௜ݑ௭ሺ௜ሻߝሺ̴ࣳܿ௜ሻ࢚࢖࢓࢕࢘࢖ࡰ࢘࢕࢚ࢇ࢔࢏࢓࢏࢘ࢉ࢙࢏ࡰ ࣳݓ Figure 2. An overview of the proposed LCP-GAN. The generator Gsynthesizes class-speciﬁc images, conditioned on the combination of intra-cluster label embeddings:/summationtext iu(i) zE(cyi). In addition to class-conditional adversarial training via the discriminator Dclass ,w e incorporate an additional discriminator Dprompt to distinguish real images from the synthesized ones, conditioned on the combination of the learnable cluster-speciﬁc prompts {ty1,...,t yky}. By competing with Dclass andDprompt ,Glearns to associate the cluster label embeddings with the underlying visual concepts described by the prompts. conditional discriminators. There have also been some at- tempts at improving GANs via unsupervised data partition- ing [ 14,16,21,31,39]. SphericGAN [ 6] imposed hard par- titioning on the training data, and aligned the real and syn-thesized data clusters in a hyper-spherical latent space. The existing semi-supervised GANs are conditioned on class i- dentity, while the semantics encapsulated in the class names is overlooked. This impedes the generative model from using prior knowledge in natural language as humans do. Furthermore, only one single label embedding is learnt per class, which is insufﬁcient to account for large intra-classvariance. To address these issues, we explore intra-classvariation factors by performing soft and ﬁner partitions on each class and learning cluster-speciﬁc prompts to represent underlying visual concepts as shown in Figure 1. More speciﬁcally, we propose a Learnable Cluster Prompt-based GAN to facilitate semi-supervised class-conditional image generation, and our model is referred to as LCP-GAN. To better match class-speciﬁc data distri-bution, the generator learns to synthesize high-ﬁdelity im-ages, conditioned on a combination of intra-class cluster la-bel embeddings. Capturing intra-class variation factors is anon-trivial task, since the semantics reﬂected by the clusters may not be well-deﬁned. Considering that natural language can express a wide range of visual concepts, we make anattempt to learn from CLIP [ 42], which provides an effec- tive way to understand the content of images. Inspired by CoOp [ 56], we can model the cluster-speciﬁc context words with learnable vectors, and further learn a mapping to adapt the CLIP representation to our generation task. As a re- sult, the generator is guided to capture cluster semantics in the adversarial training process. The framework of LCP-GAN is illustrated in Figure 2. We adopt the state-of-the- art architectures: BigGAN [ 2] and StyleGAN2-ADA [ 23], and achieve signiﬁcant improvements over them, suggest-ing that semi-supervised image generation can beneﬁt from modeling intra-class variation with the language-vision pre-training. The main contributions of this work are summarized as follows: (a) We associate the intra-class cluster label em-beddings with the cluster semantics, and the expressiveness of their combination is higher than that of a single class label embedding for capturing multiple underlying modeswith diverse visual appearances. (b) To address the issuethat the visual concepts reﬂected by the clusters may not be well deﬁned, we leverage the language-vision pre-training and represent the clusters with learnable prompts. (c) Toguide the generator to capture intra-class variation factors, the cluster prompts serve as conditional information and arejointly learnt in the adversarial training process.
Yan_CIMI4D_A_Large_Multimodal_Climbing_Motion_Dataset_Under_Human-Scene_Interactions_CVPR_2023	Abstract Motion capture is a long-standing research problem. Al- though it has been studied for decades, the majority of re- search focus on ground-based movements such as walking, sitting, dancing, etc. Off-grounded actions such as climb- ing are largely overlooked. As an important type of action in sports and firefighting field, the climbing movements is challenging to capture because of its complex back poses, intricate human-scene interactions, and difficult global lo- calization. The research community does not have an in- depth understanding of the climbing action due to the lack of specific datasets. To address this limitation, we collect CIMI4D, a large rock ClImbing MotIon dataset from 12 persons climbing 13 different climbing walls. The dataset consists of around 180,000 frames of pose inertial mea- surements, LiDAR point clouds, RGB videos, high-precision static point cloud scenes, and reconstructed scene meshes. Moreover, we frame-wise annotate touch rock holds to fa- ∗Equal contribution. †Corresponding author.cilitate a detailed exploration of human-scene interaction. The core of this dataset is a blending optimization pro- cess, which corrects for the pose as it drifts and is af- fected by the magnetic conditions. To evaluate the merit of CIMI4D, we perform four tasks which include human pose estimations (with/without scene constraints), pose pre- diction, and pose generation. The experimental results demonstrate that CIMI4D presents great challenges to ex- isting methods and enables extensive research opportuni- ties. We share the dataset with the research community in http://www.lidarhumanmotion.net/cimi4d/.	1. Introduction Capturing human motions can benefit many downstream applications, such as AR/VR, games, movies, robotics, etc. However, it is a challenging and long-standing prob- lem [1, 7,35,37,75] due to the diversity of human poses and complex interactive environment. Researchers have pro- posed various approaches to estimate human poses from images [15, 16,19,30,67], point clouds [33], inertial mea- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 12977 surement units (IMUs) [21, 69], etc. Although the prob- lem of human pose estimation (HPE) has been studied for decades [6, 58,63], most of the existing solutions focus on upright frontal poses on the ground (such as walking, sitting, jumping, dancing and yoga) [54]. Different from daily activities (such as walking and running) that are on the ground, climbing is an activity off the ground with back poses, which is also an important type for sports [22, 50], entertainment [31, 32,55], and firefighting. Climbing is an activity that involves ascending geo- graphical objects using hands and feet, such as hills, rocks, or walls. Estimating the pose of a climbing human is chal- lenging due to severe self-occlusion and the human body’s closely contact with the climbing surface. These issues are primarily caused by complex human-scene interactions. Moreover, understanding the climbing activities requires both accurate captures of the complex climbing poses and precise localization of the climber within scenes, which is especially challenging. Many pose/mesh estimation meth- ods are data-driven methods [24, 45,54,65], relying on huge climbing motion data for training networks. So a large-scale climbing dataset is necessary for the holistic understanding of human poses. Publicly available human motion datasets are mostly in upright frontal poses [2, 36,47], which are significantly different from climbing poses. Albeit some researchers collected RGBD-based climbing videos [4] or used marker-based systems [22], their data is private and the scale of dataset is very limited. To address the limitations of current datasets and boost related research, we collect a large-scale multimodal climb- ing dataset, CIMI4D, under complex human-scene inter- action, as depicted in Fig. 1. CIMI4D consists of around 180,000 frames of time-synchronized and well-annotated LiDAR point clouds, RGB videos, and IMU measurements from 12 actors climbing 13 rock-climbing walls. 12 ac- tors include professional athletes, rock climbing enthusi- asts, and beginners. In total, we collect 42 rock climb- ing motion sequences, which enable CIMI4D to cover a wide diversity of climbing behaviors. To facilitate deep un- derstanding for human-scene interactions, we also provide high-quality static point clouds using a high-precision de- vice for seven rock-climbing walls. Furthermore, we anno- tate the rock holds (holds) on climbing walls and manually label the contact information between the human body and the holds. To obtain accurate pose and global trajectory of the human body, we devise an optimization method to an- notate IMU data, as it drifts over time [10, 59] and is subject to magnetic conditions in the environment. The comprehensive annotations in CIMI4D provide the opportunity for benchmarking various 3D HPE tasks. In this work, we focus on four tasks: human pose estima- tion with or without scene constraints, human pose predic- tion and generation. To assess the effectiveness of existingmethods on these tasks, we perform both quantitative and qualitative experiments. However, most of the existing ap- proaches are unable to capture accurately the climbing ac- tion. Our experimental results demonstrate that CIMI4D presents new challenges for current computer vision algo- rithms. We hope that CIMI4D could provide more opportu- nities for a deep understanding of human-scene interactions and further benefit the digital reconstruction for both. In summary, our contributions can be listed as below: • We present the first 3D climbing motion dataset, CIMI4D, for understanding the interaction between complex human actions with scenes. CIMI4D consists of RGB videos, LiDAR point clouds, IMU measure- ments, and high-precision reconstructed scenes. • We design an annotation method which uses multiple constraints to obtain natural and smooth human poses and trajectories. • We perform an in-depth analysis of multiple methods for four tasks. CIMI4D presents a significant challenge to existing methods.
Zhang_UniDAformer_Unified_Domain_Adaptive_Panoptic_Segmentation_Transformer_via_Hierarchical_Mask_CVPR_2023	Abstract Domain adaptive panoptic segmentation aims to miti- gate data annotation challenge by leveraging off-the-shelf annotated data in one or multiple related source domains. However, existing studies employ two separate networks for instance segmentation and semantic segmentation which lead to excessive network parameters as well as compli- cated and computationally intensive training and inference processes. We design UniDAformer, a unified domain adap- tive panoptic segmentation transformer that is simple but can achieve domain adaptive instance segmentation and semantic segmentation simultaneously within a single net- work. UniDAformer introduces Hierarchical Mask Calibra- tion (HMC) that rectifies inaccurate predictions at the level of regions, superpixels and pixels via online self-training on the fly. It has three unique features: 1) it enables uni- fied domain adaptive panoptic adaptation; 2) it mitigates false predictions and improves domain adaptive panoptic segmentation effectively; 3) it is end-to-end trainable with a much simpler training and inference pipeline. Exten- sive experiments over multiple public benchmarks show that UniDAformer achieves superior domain adaptive panoptic segmentation as compared with the state-of-the-art.	1. Introduction Panoptic segmentation [30] performs instance segmenta- tion for things and semantic segmentation for stuff, which assigns each image pixel with a semantic category and a unique identity simultaneously. With the advance of deep neural networks [5, 17–19, 31, 43], panoptic segmen- tation [4, 7–9, 29, 30, 33, 35, 54, 55] has achieved very impressive performance under the supervision of plenty of densely-annotated training data. However, collecting densely-annotated panoptic data is prohibitively laborious and time-consuming [10,11,37] which has become one ma- jor constraint along this line of research. One alternative is *Equal contribution, {jingyi.zhang, jiaxing.huang }@ntu.edu.sg. †Corresponding author, [email protected]. Figure 1. Existing domain adaptive panoptic segmentation [20] adapts things and stuff separately with two isolated networks (for instance segmentation and semantic segmentation) and fuses their outputs to produce the final panoptic segmentation as in (a), leading to excessive network parameters as well as complicated and computationally intensive training and inference. Differently, UniDAformer employs a single unified network to jointly adapt things and stuff as in (b), which involves much less parameters and simplifies the training and inference pipeline greatly. to leverage off-the-shelf labeled data from one or multiple source domains . Nevertheless, the source-trained models often experience clear performance drop while applied to various target domains that usually have different data dis- tributions as compared with the source domains [20]. Domain adaptive panoptic segmentation can mitigate the inter-domain discrepancy by aligning one or multiple la- beled source domains and an unlabeled target domain [20]. To the best of our knowledge, CVRN [20] is the only work that tackles domain adaptive panoptic segmentation chal- lenges by exploiting the distinct natures of instance segmen- tation and semantic segmentation. Specifically, CVRN in- troduces cross-view regularization to guide the two segmen- tation tasks to complement and regularize each other and achieves very impressive performance. However, CVRN re- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 11227 Multi-branch Architecture Unified Architecture PSN [30] Panoptic FCN [35] MaskFormer [9] DETR [4] mSQ mRQ mPQ mSQ mRQ mPQ mSQ mRQ mPQ mSQ mRQ mPQ Supervised Setup 75.5 60.2 47.7 79.7 73.1 59.6 79.1 62.6 51.1 79.1 64.1 51.9 Adaptation Setup 59.0 27.8 20.1 47.5 19.7 15.8 56.6 19.2 16.2 56.4 21.8 18.3 Performance Drop -16.5 -32.4 -27.6 -32.2 -53.4 -43.8 -22.5 -43.4 -34.9 -22.7 -42.3 -33.6 Table 1. Panoptic segmentation with traditional multi-branch architecture [30] and recent unified architectures [4, 9, 35]: The Supervised Setup trains with the Cityscapes [10] and tests on the same dataset. The UDA Setup trains with the SYNTHIA [45] and tests on Cityscapes. It can be seen that the performance drops between the two learning setups come more from mRQ than from mSQ consistently across different architectures. In addition, such a phenomenon is more severe for unified architectures. This demonstrates a clear false prediction issue in unified domain adaptive panoptic segmentation as mRQ is computed with false positives and false negatives. lies on a multi-branch architecture that adopts a two-phase pipeline with two separate networks as illustrated in Fig. 1 (a). This sophisticated design directly doubles network pa- rameters, slows down the training, and hinders it from being end-to-end trainable. It is desirable to have a unified panop- tic adaptation network that can effectively handle the two segmentation tasks with a single network. We design a unified domain adaptive panoptic segmenta- tion transformer (UniDAformer) as shown in Fig. 1 (b). Our design is based on the observation that one major issue in unified panoptic adaptation comes from a severe false pre- diction problem. As shown in Table 1, most recent unified panoptic segmentation architectures [4, 9, 35] outperform traditional multi-branch architectures [30] by large margins under the supervised setup. However, the situation inverts completely under unsupervised domain adaptation setup. Such contradictory results are more severe for the recog- nition quality in mRQ. This shows that the panoptic quality drop of unified architecture mainly comes from False Posi- tives (FP) and False Negatives (FN) as mRQ is computed from all predictions (True Positives, False Negatives and False Positives) while the segmentation quality in mSQ is computed with True Positives (TP) only. In the proposed UniDAformer, we mitigate the false prediction issue by introducing Hierarchical Mask Cali- bration (HMC) that calibrates inaccurate predictions at the level of regions, superpixels, and pixels. With the cor- rected masks, UniDAformer re-trains the network via an online self-training process on the fly. Specifically, HMC treats both things and stuff predictions as masks uniformly and corrects each predicted pseudo mask hierarchically in a coarse-to-fine manner, i.e., from region level that cali- brates the overall category of each mask to superpixel and pixel levels that calibrate the superpixel and pixels around the boundary of each mask (which are more susceptible to prediction errors). UniDAformer has three unique fea- tures. First , it achieves unified panoptic adaptation by treat- ing things and stuff as masks and adapting them uniformly. Second , it mitigates the false prediction issue effectivelyby calibrating the predicted pseudo masks iteratively and progressively. Third , it is end-to-end trainable with much less parameters and simpler training and inference pipeline. Besides, HMC introduces little computation overhead and could be used as a plug-in. The contributions of this work can be summarized in three aspects. First, we propose UniDAformer that en- ables concurrent domain adaptive instance segmentation and semantic segmentation within a single network. It is the first end-to-end unified domain adaptive panoptic seg- mentation transformer to the best our knowledge. Sec- ond, we design Hierarchical Mask Calibration with online self-training, which allows to calibrate the predicted pseudo masks on the fly during self-training. Third, extensive ex- periments over multiple public benchmarks show that the proposed UniDAformer achieves superior segmentation ac- curacy and efficiency as compared with the state-of-the-art.
Zhao_MetaFusion_Infrared_and_Visible_Image_Fusion_via_Meta-Feature_Embedding_From_CVPR_2023	Abstract Fusing infrared and visible images can provide more tex- ture details for subsequent object detection task. Converse- ly, detection task furnishes object semantic information to improve the infrared and visible image fusion. Thus, a joint fusion and detection learning to use their mutual promotion is attracting more attention. However, the feature gap be- tween these two different-level tasks hinders the progress. Addressing this issue, this paper proposes an infrared and visible image fusion via meta-feature embedding from ob- ject detection. The core idea is that meta-feature embedding model is designed to generate object semantic features ac- cording to fusion network ability, and thus the semantic fea- tures are naturally compatible with fusion features. It is op- timized by simulating a meta learning. Moreover, we further implement a mutual promotion learning between fusion and detection tasks to improve their performances. Comprehen- sive experiments on three public datasets demonstrate the effectiveness of our method. Code and model are available at: https://github.com/wdzhao123/MetaFusion.	1. Introduction Multi-modality sensor technology has promoted the ap- plication of multi-modality images in different areas. A- mong them, infrared images and visible images have been utilized commonly, as the information contained in these two modalities is complementary. Speciﬁcally, infrared im- ages can supply object thermal structures without being af- fected by illumination. But they are short of texture de- tails. On the contrary, visible images can catch the texture information for the scene. But they are severely affected by light. Thus, many methods [15,25,35,43–45,47] focus on s- Corresponding author (a) Separate optimization method (b) Cascaded optimization method (c) Meta-feature embedding method IVIF Network OD Network min ((,;)) ff f L xy  min((;)) dd d Lz  IVIF Network OD Network min ((,;)) ((;)) ff f d d L xy Lz    (,;)f z xy  IVIF Network OD Network(,;)f z xy  Meta-feature Embeddingmin ((,;)) (,(,;)) ff f d f f gi i im Lxy LFFF     min ((,;)) mf f Lxy       Data flow Gradient flow(,;)f z xy Figure 1. Different joint learning methods of infrared and visible image fusion (IVIF) and object detection (OD). (a) Separate opti- mization method: IVIF network is ﬁrstly optimized by fusion lossLf. Then, fusion result zis generated by from the input infrared and visible image pair x; y. Finally, OD network is op- timized by detection loss Ldusing z. (b) Cascaded optimization method: OD network is treated as a constraint to optimize IVIF network by the loss LfandLd. (c) Meta-feature embedding method: Meta-feature embedding is optimized to learn how to guide to have a low Lf. Then, generates meta feature from detection feature Fd iand fusion feature Fd i. Finally, the meta fea- ture is used to guide by the loss Ld. tudying pixel-level infrared and visible image fusion (IVIF), thereby helping high-level tasks improve performance, e.g., object detection (OD) [20, 34]. IVIF and OD can greatly beneﬁt form each other. IVIF This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 13955 generates the fused image that contains more information than any single modality image to improve OD. Howev- er, IVIF mainly focuses on the pixel relationship between an image pair, and there is little consideration for object semantic. In contrast, OD can provide rich object seman- tic information to IVIF, as its aim is locating the objects. Therefore, this paper studies a joint learning framework be- tween IVIF and OD to improve their performances. Existing joint learning methods of IVIF and OD can be divided into two categories: separate optimization and cascaded optimization. Separate optimization ﬁrstly train- s IVIF network, and then trains OD network using IVIF results, as shown in Figure 1(a). Thus, most methods fo- cus on improving the fusion effect, e.g., designing network- s [13, 23, 35, 36] and introducing constraints [10, 26, 45]. Obviously, separate optimization neglects the help of OD. Cascaded optimization adopts OD network as a constraint to train IVIF network, and thus forces the IVIF network to generate fusion images with easily detected objects [20], as shown in Figure 1(b). However, directly utilizing the high- level OD constraint to guide the pixel-level IVIF will result in limited effect. Therefore, we leverage OD feature maps that guide IVIF feature maps to obtain more semantic infor- mation. Unfortunately, OD features are mismatched with IVIF features due to their task-level difference. Addressing this issue, we propose a meta-feature embedding network (MFE ), as shown in Figure 1(c). The idea is that if MFE generates OD features according to the IVIF network abil- ity, the OD features are naturally compatible with the IVIF network, and the optimization can be achieved by simulat- ing a meta learning. Speciﬁcally, an infrared and visible image fusion via meta-feature embedding from object detection is proposed, which is named as MetaFusion. MetaFusion includes IVIF network (F), OD network ( D), andMFE . In particular, MFE is expected to generate meta features to bridge the gap between FandD, which is optimized by two alter- nate steps: inner update and outer update. In the inner up- date process, we ﬁrstly optimize Fusing meta training set Smtrto obtain its updated network F0. Then,F0calculates the fusion loss on meta testing set Smtsto optimizeMFE . The motivation is that if MFE successfully generates meta features which are compatible with F,F0will produce bet- ter fused images, i.e., the fusion loss should be lower. In the outer update process, Fis optimized with the guide of the meta features generated by the ﬁxed MFE onSmtr andSmts. In this way, Fcan learn how to extract seman- tic information to improve fusion quality. In the above two alternate steps, Dis ﬁxed to offer detection semantic infor- mation. Thus, we further implement a mutual promotion learning, where we use the improved Fto generate fusion results to ﬁnetune D, and then the improved Doffers better semantic information to optimize F.In summary, our contributions are as follows. (1) We ex- plore the joint learning framework of IVIF and OD, and pro- pose MetaFuison to obtain superior performance on these two tasks. (2) Meta-feature embedding network is designed to generate meta features that bridge the gap between Fand D. (3) Sequentially, mutual promotion learning between F andDis introduced to improve their performances. (4) Ex- tensive experiments on image fusion and object detection validate the effectiveness of the proposed method.
Zheng_HS-Pose_Hybrid_Scope_Feature_Extraction_for_Category-Level_Object_Pose_Estimation_CVPR_2023	Abstract In this paper, we focus on the problem of category-level object pose estimation, which is challenging due to the large intra-category shape variation. 3D graph convolution (3D- GC) based methods have been widely used to extract local geometric features, but they have limitations for complex shaped objects and are sensitive to noise. Moreover, the scale and translation invariant properties of 3D-GC restrict the perception of an object’s size and translation informa- tion. In this paper, we propose a simple network structure, theHS-layer , which extends 3D-GC to extract hybrid scope latent features from point cloud data for category-level ob- ject pose estimation tasks. The proposed HS-layer: 1) is able to perceive local-global geometric structure and global information, 2) is robust to noise, and 3) can encode size and translation information. Our experiments show that the simple replacement of the 3D-GC layer with the proposed HS-layer on the baseline method (GPV-Pose) achieves a significant improvement, with the performance increased by 14.5% on5◦2cm metric and 10.3% on IoU 75. Our method outperforms the state-of-the-art methods by a large margin (8.3% on5◦2cm,6.9% on IoU 75) on REAL275 dataset and runs in real-time (50 FPS)1.	1. Introduction Accurate and efficient estimation of an object’s pose and size is crucial for many real-world applications [48], in- cluding robotic manipulation [15], augmented reality [36], and autonomous driving, among others. In these appli- *The corresponding author. 1Code is available: https://github.com/Lynne-Zheng-Linfang/HS-Pose Input Point Cloud 6D Pose and 3D Size EstimationHybrid Scope Feature Extraction Scale and Translation Encoding Local Geometric Structure xyzxz �� ≠ xGlobal Geometric StructureOutlier Robust Global Information y �, �, � ��Figure 1. Illustration of the hybrid scope feature extraction of the HS-layer . As shown in the right figure, the proposed HS-layer possesses various advantages, including the capability of capturing both local and global geometric information, robustness to outliers, and the encoding of scale and translation information. Building upon the GPV-pose, the HS- layer is employed to develop a category-level pose estimation framework, namely HS-Pose . Upon receiving an input point cloud, HS-Pose outputs the estimated 6D pose and 3D size of the object, as shown in the left fig- ure. Given the strengths of the HS-layer, HS-Pose is capable of handling complex object shapes, exhibits robustness to outliers, and achieves better performance compared with existing methods. cations, it is essential that pose estimation algorithms can handle the diverse range of objects encountered in daily life. While many existing works [3,13,29,50] have demon- strated impressive performance in estimating an object’s pose, they typically focus on only a limited set of ob- jects with known shapes and textures, aided by CAD mod- els. In contrast, category-level object pose estimation al- gorithms [7, 22, 34, 45, 49] address all objects within a given category and enable pose estimation of unseen objects during inference without the target objects’ CAD models, This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 17163 which is more suitable for daily-life applications. However, developing such algorithms is more challenging due to the shape and texture diversity within each category. In recent years, category-level object pose estimation re- search [5, 55, 56] has advanced rapidly by adopting state- of-the-art deep learning methods. [2, 46] gain the ability to generalize by mapping the input shape to normalized or metric-scale canonical spaces and then recovering the ob- jects’ poses via correspondence matching. Better handling of intra-category shape variation is also achieved by lever- aging shape priors [4, 42, 56], symmetry priors [20], or do- main adaptation [17,21]. Additionally, [5] enhances the per- ceptiveness of local geometry, and [7,55] exploit geometric consistency terms to improve the performance further. Despite the remarkable progress of existing methods, there is still room for improvement in the performance of the category-level object pose estimation. Reconstruction and matching-based methods [17, 42, 46] are usually lim- ited in speed due to the time-consuming correspondence- matching procedure. Recently, various methods [5, 7, 20, 55, 56] built on 3D graph convolution (3D-GC) [23] have achieved impressive performance and run in real-time. They show outstanding local geometric sensitivity and the ability to generalize to unseen objects. However, only look- ing at small local regions impedes their ability to leverage the global geometric relationships that are essential for han- dling complex geometric shapes and makes them vulnerable to outliers. In addition, the scale and translation invariant nature of 3D-GC restrict the perception of object size and translation information. To overcome the limitations of 3D-GC in category-level object pose estimation, we propose the hybrid scope latent feature extraction layer (HS-layer), which can perceive both local and global geometric relationships and has a better awareness of translation and scale. Moreover, the proposed HS-layer is highly robust to outliers. To demonstrate the ef- fectiveness of the HS-layer, we replace the 3D-GC layers in GPV-Pose [7] to construct a new category-level object pose estimation framework, HS-pose. This framework signifi- cantly outperforms the state-of-the-art method and runs in real time. Our approach extends the perception of 3D-GC to incorporate other essential information by using two paral- lel paths for information extraction. The first path encodes size and translation information (STE), which is missing in 3D-GC due to its invariance property. The second path ex- tracts outlier-robust geometric features using the receptive field with the feature distance metric (RF-F) and the outlier- robust feature extraction layer (ORL). The main contribution of this paper is as follows: • We propose a network architecture, the hybrid scope latent feature extraction layer (HS-layer), that can simultaneously perceive local and global geometric structure, encode translation and scale information,and extract outlier-robust feature information. Our proposed HS-layer balances all these critical aspects necessary for category-level pose estimation. • We use the HS-layer to develop a category-level pose estimation framework, HS-Pose, based on GPV-Pose. The HS-Pose, when compared to its parent frame- work, has an advantage in handling complex geometric shapes, capturing object size and translation while be- ing robust to noise. • We conduct extensive experiments and show that the proposed method can handle complex shapes and out- performs the state-of-the-art methods by a large mar- gin while running in real-time (50FPS).
Zhu_R2Former_Unified_Retrieval_and_Reranking_Transformer_for_Place_Recognition_CVPR_2023	Abstract Visual Place Recognition (VPR) estimates the location of query images by matching them with images in a reference database. Conventional methods generally adopt aggre- gated CNN features for global retrieval and RANSAC-based geometric verification for reranking. However, RANSAC only employs geometric information but ignores other pos- sible information that could be useful for reranking, e.g. lo- cal feature correlations, and attention values. In this paper, we propose a unified place recognition framework that han- dles both retrieval and reranking with a novel transformer model, named R2Former. The proposed reranking module takes feature correlation, attention value, and xy coordi- nates into account, and learns to determine whether the image pair is from the same location. The whole pipeline is end-to-end trainable and the reranking module alone can also be adopted on other CNN or transformer back- bones as a generic component. Remarkably, R2Former significantly outperforms state-of-the-art methods on ma- jor VPR datasets with much less inference time and mem- ory consumption. It also achieves the state-of-the-art on the hold-out MSLS challenge set and could serve as a sim- ple yet strong solution for real-world large-scale applica- tions. Experiments also show vision transformer tokens are comparable and sometimes better than CNN local fea- tures on local matching. The code is released at https: //github.com/Jeff-Zilence/R2Former .	1. Introduction Visual Place Recognition (VPR) aims to localize query images from unknown locations by matching them with a set of reference images from known locations. It has great potential for robotics [47], navigation [41], autonomous driving [9], augmented reality (AR) applications. Previ- ous works [26, 53] generally formulate VPR as a retrieval problem with two stages, i.e. global retrieval, and rerank- ing. The global retrieval stage usually applies aggregation †This work was done during the first author’s internship at ByteDance CNN Local Features Number of Inliers (b) Our Unified Framework (a) Conventional PipelineNetVLAD (Retrieval ) Transformers ( Retrieval ) Reranking ScoreTransformers ( Reranking )Selected Patch Pairs…… (c) Comparison with State -of-the-art MethodsRANSAC ( Reranking ) Transformer Tokens 40506070 MSLS Challenge (R@1)Patch-NetVLAD [26] TransVPR [53] Ours00.511.5 Infer. Time (s)02040 Memory (GB)+9.1 4.7× Reduction4.7× ReductionFigure 1. An overview of conventional pipeline and our unified framework. Both our global retrieval and reranking modules are end-to-end trainable transformer layers, which achieve state-of- the-art performance with much less computational cost. methods ( e.g. NetVLAD [3] and GeM [43]) on top of CNN (Convolutional Neural Network) to retrieve top candidates from a large reference database. While some works [3, 6] only adopt global retrieval, current state-of-the-art meth- ods [26, 53] conduct reranking ( i.e. geometric verification with RANSAC [19]) on the top-k ( e.g.k= 100 ) candidates to further confirm the matches, typically leading to a signif- icant performance boost. However, geometric information is not the only information that could be useful for rerank- ing, and task-relevant information could be learned with a data-driven module to further boost performance. Besides, the current reranking process requires a relatively large in- ference time and memory footprint (typically over 1s and 1MB per image), which cannot scale to real-world appli- cations with large QPS (Queries Per Second) and reference database ( >1M images). Recently vision transformer [17] has achieved signifi- cant performance on a wide range of vision tasks, and has been shown to have a high potential for VPR [53] (Sec. 2). This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 19370 However, the predominant local features [26, 53] are still based on CNN backbones, due to the built-in feature lo- cality with limited receptive fields. Although vision trans- former [17] considers each patch as an input token and natu- rally encodes local information, the local information might be overwritten with global information by the strong global correlation between all tokens in every layer. Therefore, it is still unclear how vision transformer tokens perform in local matching as compared to CNN local features in this field. In this paper, we integrate the global retrieval and rerank- ing into one unified framework employing only transform- ers (Fig. 1), abbreviated as R2Former, which is simple, ef- ficient, and effective. The global retrieval is tackled based on the class token without additional aggregation modules [3, 43] and the other image tokens are adopted as local fea- tures. Different from geometric verification which focuses on geometric information between local feature pairs, we feed the correlation between the tokens (local features) of a pair of images, xy coordinates of tokens, and their atten- tion information to transformer modules, so that the mod- ule can learn task-relevant information that could be use- ful for reranking. The global retrieval and reranking parts can be either trained in an end-to-end manner or tuned al- ternatively with a more stable convergence. The proposed reranking module can also be adopted on other CNN or transformer backbones, and our comparison shows that vi- sion transformer tokens are comparable to CNN local fea- tures in terms of reranking performance (Table 6). Without bells and whistles, the proposed method out- performs both retrieval-only and retrieval+reranking state- of-the-art methods on a wide range of VPR datasets. The proposed method follows a very efficient design and ap- plies linear layers for dimension reduction: i.e. only 256and 500×131for global and local feature dimensions and only 32for transformer dimension of reranking module, thus is significantly faster ( >4.7×QPS) with much less ( <22%) memory consumption than previous methods [26,53]. Both the global and local features are extracted from the same backbone model only once, and the reranking of top-k can- didates is finished with only one forward pass by computing the reranking scores of all candidate pairs in parallel within one batch. The reranking speed can be further boosted by parallel computing on multiple GPUs with >20×speedup over previous methods [26, 53]. We demonstrate that the proposed reranking module also learns to focus on good lo- cal matches like RANSAC [19]. We summarize our contri- butions as follows: • A unified retrieval and reranking framework for place recognition employing pure transformers, which demon- strates that vision transformer tokens are comparable and sometimes better than CNN local features in terms of reranking or local matching. • A novel transformer-based reranking module that learnsto attend to the correlation of informative local feature pairs. It can be combined with either CNN or transformer backbones with better performance and efficiency than other reranking methods, e.g. RANSAC. • Extensive experiments showing state-of-the-art perfor- mance on a wide range of place recognition datasets with significantly less ( <22%) inference latency and memory consumption than previous reranking-based methods.
Zhao_OmniAL_A_Unified_CNN_Framework_for_Unsupervised_Anomaly_Localization_CVPR_2023	Abstract Unsupervised anomaly localization and detection is cru- cial for industrial manufacturing processes due to the lack of anomalous samples. Recent unsupervised advances on industrial anomaly detection achieve high performance by training separate models for many different categories. The model storage and training time cost of this paradigm is high. Moreover, the setting of one-model-N-classes leads to fearful degradation of existing methods. In this pa- per, we propose a uniﬁed CNN framework for unsuper- vised anomaly localization, named OmniAL. This method conquers aforementioned problems by improving anomaly synthesis, reconstruction and localization. To prevent the model learning identical reconstruction, it trains the model with proposed panel-guided synthetic anomaly data rather than directly using normal data. It increases anomaly re- construction error for multi-class distribution by using a network that is equipped with proposed Dilated Channel and Spatial Attention (DCSA) blocks. To better localize the anomaly regions, it employs proposed DiffNeck between reconstruction and localization sub-networks to explore multi-level differences. Experiments on 15-class MVTecAD and 12-class VisA datasets verify the advantage of proposed OmniAL that surpasses the state-of-the-art of uniﬁed mod- els. On 15-class-MVTecAD/12-class-VisA, its single uniﬁedmodel achieves 97.2/87.8 image-AUROC, 98.3/96.6 pixel- AUROC and 73.4/41.7 pixel-AP for anomaly detection and localization respectively. Besides that, we make the ﬁrst at- tempt to conduct a comprehensive study on the robustness of unsupervised anomaly localization and detection meth- ods against different level adversarial attacks. Experiential results show OmniAL has good application prospects for its superior performance.	1. Introduction In real industrial scenarios, the location of anomaly [22, 26] reveals important information, such as defective types and degrees. It is essential not only to inspect whether a sample is defective but also to know where the speciﬁc anomaly regions are. Since anomaly appearance is inex- haustible, it is almost impossible and infeasible to collect and manually annotate all kinds of abnormal data. Thus, only normal samples are available for training a detector that is robust enough to ﬁnd out unseen anomalies during in- ference phase. Considering the diversity of classes and var- ious types of one class, the conventional training paradigm of N models for N classes, as shown in Fig.1a, may not be the best solution. The model storage and training time cost increase with the number of classes. As shown in Fig.1b, existing method severely degrades anomaly local- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 3924 Figure 2. Problem analysis. The ﬁnal failure may caused by reconstruction and localization. ization performance if the training paradigm changes to one model for N classes. Therefore, a robust uniﬁed framework for unsupervised anomaly localization is highly demanded for intelligent industrial. With the limitation of available training data, many ap- pealing unsupervised approaches [14, 18, 35, 38] using syn- thesized anomaly data are proposed. These approaches gen- erate anomalous instances to inspire the anomaly detector to learn discriminative features. Their experiments show that the realisticness of generated anomalous instances had a strong impact on the quality of anomaly localization. How- ever, none of these methods consider the training paradigm of one model for N classes. When switching to the uniﬁed training paradigm, they are more prone to learn an identical short-cut and fail to discriminate the anomaly. With the normal and synthesized anomalous sam- ples, recent unsupervised learning methods train a deep anomaly detector by either a distance-based [6, 14, 19–21, 27] or reconstruction-based [2, 9, 35, 36, 38] way. The reconstruction-based architectures [35, 38] are supposed to reconstruct normal images more accurately than the un- seen anomalous. The anomaly localization is then cal- culated from the reconstruction error between the original and reconstructed versions of the input image, as shown in Fig.2a. The prediction of anomaly location is not only based on the reconstruction quality but also the ability of spot- ting the reconstruction error. The typical reconstruction- based method JNLD [38] learns a joint representation of an anomalous image and its anomaly-free reconstruction, while simultaneously learning a decision boundary between normal and simulated anomalous examples. As shown in Fig.2b and Fig.2c, under the uniﬁed setting, JNLD [38] fails to produce correct results either because of the reconstruc- tion failure or the localization failure. To conquer aforementioned problems, we propose a novel uniﬁed framework OmniAL for effectively localiz-ing anomaly pixels of different classes only by using a sin- gle model. OmniAL uses a panel-guided anomaly synthe- sis method that controls the portion of normal and anomaly regions for each training sample. By doing this, OmniAL blocks the chance of learning identical shortcut from the source. To increase the anomaly reconstruction error for multi-class distribution, OmniAL constructs a reconstruc- tion and a localization sub-networks that are equipped with proposed Dilated Channel and Spatial Attention (DCSA) blocks. To better localize the anomaly regions, OmniAL employs a DiffNeck module between the reconstruction and localization sub-networks to explore multi-level reconstruc- tion errors. As shown in Fig.1 and Fig.2, OmniAL learns a single uniﬁed model for multiple classes that produces high quality reconstruction and precise anomaly localiza- tion. Furthermore, we conduct an exhaustive evaluation of reconstruction and localization performance against to multi-level adversarial attacks. In summary, we make following main contributions: We construct a uniﬁed CNN framework OmniAL for unsupervised anomaly localization that is equipped with proposed panel-guided anomaly synthesis, DCSA block, and DiffNeck module. OmniAL achieves supe- rior performance for anomaly localization on challeng- ing MVTecAD [1] and VisA [39] datasets compared to the state-of-the-art. By preventing model from learning identical re- construction, our proposed panel-guided anomaly synthesis method also brings substantial improve- ment for existing methods under the uniﬁed setting. It boosts the image-AUROC/pixel-AUROC/pixel-AP from 88.7/87.1/49.4 to 92.5/94.5/57.4 for Draem [35]. We make a comprehensive study on the robustness of separate/uniﬁed anomaly localization methods against different level adversarial attacks. Our synthesized 3925 Figure 3. Framework of OmniAL. It consists of panel-guided anomaly synthesis, reconstruction and localization. Anomaly synthesis is based on anomaly panel and three variants of the Just Noticeable Distortion (JND) map. The synthetic anomaly is reconstructed into normal image and corresponding JND map by the Dilated Channel Spatial Attention (DCSA) modules equipped reconstruction sub-network. The localization sub-network with a DiffNeck module localizes the anomaly regions by exploring the difference between reconstructed and original data. adversarial datasets exhibit strong attack capability against anomaly detection, reconstruction and local- ization, also helping to analyse the risks of existing methods.
Zheng_Prototype-Based_Embedding_Network_for_Scene_Graph_Generation_CVPR_2023	Abstract Current Scene Graph Generation (SGG) methods ex- plore contextual information to predict relationships among entity pairs. However, due to the diverse visual appearance of numerous possible subject-object combinations, there is a large intra-class variation within each predicate cat- egory, e.g., “man-eating -pizza, giraffe-eating -leaf”, and the severe inter-class similarity between different classes, e.g., “man-holding -plate, man-eating -pizza”, in model’s la- tent space. The above challenges prevent current SGG methods from acquiring robust features for reliable rela- tion prediction. In this paper, we claim that the predi- cate’s category-inherent semantics can serve as class-wise prototypes in the semantic space for relieving the chal- lenges. To the end, we propose the Prototype-based Embed- ding Network (PE-Net) , which models entities/predicates with prototype-aligned compact and distinctive representa- tions and thereby establishes matching between entity pairs and predicates in a common embedding space for relation recognition. Moreover, Prototype-guided Learning (PL) is introduced to help PE-Net efficiently learn such entity- predicate matching, and Prototype Regularization (PR) is devised to relieve the ambiguous entity-predicate match- ing caused by the predicate’s semantic overlap. Exten- sive experiments demonstrate that our method gains su- perior relation recognition capability on SGG, achieving new state-of-the-art performances on both Visual Genome and Open Images datasets. The codes are available at https://github.com/VL-Group/PENET .	1. Introduction Scene Graph Generation (SGG) is a fundamental com- puter vision task that involves detecting the entities and predicting their relationships in an image to generate a scene graph, where nodes indicate entities and edges in- *Equal contribution. †Corresponding author. (a)man-holding-plate (b)person-holding-rocket (c)man-eating-pizza (e)giraffe-eating-leaf (d)man-eating-pizzaFigure 1. The illustration of relation representations with large intra-class variation and severe inter-class similarity. Left: the feature distribution of “eating” (in red) and “holding” (in blue) obtained by Motifs [39]. Right: some instances of “eating” and “holding”. Examples (c) and (e) illustrate that relation instances from the same class have diverse appearance. Moreover, examples (a) and (c) demonstrate that similar-looking relation instances may belong to different categories. dicate relationships between entity pairs. Such a graph- structured representation is helpful for downstream tasks such as Visual Question Answering [5, 15, 41], Image Cap- tioning [4, 38, 42, 45], and Image Retrieval [10, 25, 40]. Existing SGG models [1, 3, 7, 13, 18, 29, 39] typically start with an object detector that generates a set of entity proposals and corresponding features. Then, entity fea- tures are enhanced by exploring the contextual informa- tion taking advantage of message-passing modules. Fi- nally, these refined entity features are used to predict pair- wise relations. Although many works have made great ef- forts to explore the contextual information for robust re- lation recognition, they still suffer from biased-prediction problems, preferring common predicates ( e.g.,“on”, “of”) instead of fine-grained ones ( e.g.,“walking on”, “cover- ing”). To address the problem, various de-biasing frame- works [6, 14, 21, 28, 34, 37, 46, 47] have been proposed to obtain balanced prediction results. While alleviating the long-tailed issue to some extent, most of them only achieve a trade-off between head and tail predicates. In other words, they sacrifice the robust representations learned on head This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 22783 predicates for unworthy improvements in the tail ones [46], which do not truly improve the model’s holistic recognition ability for most of the relations. The origin of the issue lies in the fact that current SGG methods fail to capture compact and distinctive represen- tations for relations. For instance, as shown in Fig. 1, the relation representation, derived from Motifs’ latent space, is heavily discrete and intersecting. Hence, it makes exist- ing SGG models hard to learn perfect decision boundaries for accurate predicate recognition. Accordingly, we sum- marize the issue as two challenges: large Intra-class vari- ation within the same relation class and severe Inter-class similarity between different categories. Intra-class variation. The intra-class variation arises from the diverse appearance of entities and various subject- object combinations. Specifically, entities’ visual appear- ances change greatly even though they belong to the same class. Thus, represented as the union feature containing subject and object entities, relation representations signif- icantly vary with the appearances of entity instances, e.g., various visual representations for “pizza” in Fig. 1(c) vs. Fig. 1(d). Besides, the numerous subject-object combina- tions of predicate instances further increase the variation within each predicate class, e.g., “man-eating-pizza” vs. “giraffe-eating-leaf” in Fig. 1(c) and Fig. 1(e). Inter-class similarity. The inter-class similarity of rela- tions originates from similar-looking interactions but be- longs to different predicate classes. For instance, as shown in Fig. 1(a) and Fig. 1(c), the similar visual appearance of interactions between “man-pizza” and “man-plate” make current SGG models hard to distinguish “eating” from “holding”, even if they are semantic irrelevant to each other. The above challenges motivate us to study two problems: 1) For the intra-class variation, how to capture category- inherent features, producing compact representations for entity/predicate instances from the same category. More- over, 2) for the inter-class similarity, how to derive distinc- tive representations for effectively distinguishing similar- looking relation instances between different classes. Our key intuition is that semantics is more reliable than visual appearance when modeling entities/predicates. Intuitively, although entities/predicates of the same class significantly vary in visual appearance, they all share the representa- tive semantics, which can be easily captured from their class labels. Dominated by the representative semantics, the representations of entities and predicates have smaller variations within their classes in the semantic space. Be- sides, the class-inherent semantics is discriminative enough for visual-similar instances between different categories. Therefore, in conjunction with the above analysis, model- ing entities and predicates in the semantic space can provide highly compact and distinguishable representations against intra-class variation and inter-class similarity challenges.Inspired by that, we propose a simple but effective method, Prototype-based Embedding Network (PE-Net), which produces compact and distinctive entity/predicate representations for relation recognition. To achieve that, the PE-Net models entity and predicate instances with com- pact and distinguishable representations in the semantic space, which are closely aligned to their semantic proto- types. Practically, the prototype is defined as the represen- tative embedding for a group of instances from the same entity/predicate class. Then, the PE-Net establishes match- ing between entity pairs ( i.e., subject-object ( s,o)) and their corresponding predicates ( p) for relation recognition ( i.e., F(s,o)≈p). Besides, a Prototype-guided Learning strat- egy (PL) is proposed to help PE-Net efficiently learn this entity-predicate matching. Additionally, to alleviate the am- biguous entity-predicate matching caused by the semantic overlap between predicates ( e.g., “walking on” and “stand- ing on”), Prototype Regularization (PR) is proposed to en- courage inter-class separation between predicate prototypes for precise entity-predicate matching. Finally, we introduce two metrics, i.e., Intra-class Variance (IV) and Intra-class to Inter-class Variance Ratio (IIVR), to measure the compact- ness and distinctiveness of entity/predicate representations, respectively. In summary, the main contributions of our work are three folds: • We propose a simple yet effective method, i.e., Prototype-based Embedding Network (PE-Net), which produces compact and distinctive entity/predicate rep- resentations and then establishes matching between entity pairs and predicates for relation recognition. • Moreover, Prototype-guided Learning (PL) is intro- duced to help PE-Net efficiently learn such entity- predicate matching, and Prototype Regularization (PR) is devised to relieve the ambiguous entity-predicate matching caused by the predicate’s semantic overlap. • Evaluated on the Visual Genome and Open Images datasets, our method significantly increases the rela- tion recognition ability for SGG, achieving new state- of-the-art performances.
Zhou_Shifted_Diffusion_for_Text-to-Image_Generation_CVPR_2023	Abstract We present Corgi, a novel method for text-to-image gen- eration. Corgi is based on our proposed shifted diffusion model, which achieves better image embedding generation from input text. Unlike the baseline diffusion model used in DALL-E 2, our method seamlessly encodes prior knowledge of the pre-trained CLIP model in its diffusion process by designing a new initialization distribution and a new tran- sition step of the diffusion. Compared to the strong DALL-E 2 baseline, our method performs better in generating im- age embedding from the text in terms of both efficiency and effectiveness, resulting in better text-to-image genera- tion. Extensive large-scale experiments are conducted and evaluated in terms of both quantitative measures and hu- man evaluation, indicating a stronger generation ability of our method compared to existing ones. Furthermore, our model enables semi-supervised and language-free training for text-to-image generation, where only part or none of the images in the training dataset have an associated caption. Trained with only 1.7%of the images being captioned, our semi-supervised model obtains FID results comparable to DALL-E 2 on zero-shot text-to-image generation evaluated on MS-COCO. Corgi also achieves new state-of-the-art re- sults across different datasets on downstream language-free text-to-image generation tasks, outperforming the previous method, Lafite, by a large margin.	1. Introduction “AI-generated content” has attracted increasingly more public awareness thanks to the significant progress in re- cent research of high-fidelity text-aligned image synthetic tasks. [7, 20–23, 26, 32]. Particularly, models trained on web-scale datasets have demonstrated their impressive abil- ity to generate out-of-distribution images from arbitrary text inputs that describe unseen combinations of visual con- *Performed this work during internship at ByteDance, code is available athttps://github.com/drboog/Shifted_Diffusion . The research of the first and last author was supported in part by NSF through grants IIS-1910492 and CCF-2200173 and by KAUST CRG10-4663.2.cepts. Starting from DALL-E [21], researchers have proposed a variety of approaches to further advance the state-of-the-art (SOTA) of text-to-image generation in terms of both gen- eration quality and efficiency. Latent Diffusion Model [22] trains a diffusion model in the latent space of auto-encoder instead of pixel space, leading to better generation effi- ciency. GLIDE [15] adopts a hierarchical architecture, which consists of diffusion models at different resolutions. Such a model design strategy has shown to be effective and has been adopted by many follow-up works. DALL-E 2 [20] further introduces an extra image embedding input. Such an image embedding not only improves the model per- formance in text-to-image generation but also enables ap- plications, including image-to-image generation and gener- ation under multi-modal conditions. Imagen [23] makes use of a rich pre-trained text encoder [19], demonstrating that a frozen text encoder pre-trained on the large-scale text-only dataset can help text-to-image generation models in under- standing the semantics of text descriptions. Parti [32] shows a further successful scale-up of the generative model, lead- ing to impressive improvement in text-to-image consistency with transformer structure. The aforementioned approaches focus on improv- ing text-to-image generation by either scaling up train- able/frozen modules or designing better model architec- tures. In this work, we explore an orthogonal direction, where we propose novel techniques to improve the diffu- sion process itself and make it more suitable and effective for text-to-image generation. Specifically, we propose Corgi ( Clip-based shifted dif- fusiOn model b RidGIng the gap), a novel diffusion model designed for a flexible text-to-image generation. Our model can perform text-to-image generation under all the super- vised, semi-supervised, and language-free settings. By “bridging the gap,” we emphasize two key novelties in our method: (1) our model tries to bridge the image-text modal- ity gap [12] so as to train a better generative model. Modal- ity gap is a critical concept discovered in pre-trained mod- els such as CLIP [17], which captures the phenomenon that multi-modality representations do not align in the joint em- This CVPR paper is the Open Access v
Zhou_RepMode_Learning_to_Re-Parameterize_Diverse_Experts_for_Subcellular_Structure_Prediction_CVPR_2023	Abstract In biological research, fluorescence staining is a key technique to reveal the locations and morphology of subcel- lular structures. However, it is slow, expensive, and harm- ful to cells. In this paper, we model it as a deep learning task termed subcellular structure prediction (SSP), aiming to predict the 3D fluorescent images of multiple subcellu- lar structures from a 3D transmitted-light image. Unfortu- nately, due to the limitations of current biotechnology, each image is partially labeled in SSP . Besides, naturally, sub- cellular structures vary considerably in size, which causes the multi-scale issue of SSP . To overcome these challenges, we propose Re-parameterizing Mixture-of-Diverse-Experts (RepMode), a network that dynamically organizes its pa- rameters with task-aware priors to handle specified single- label prediction tasks. In RepMode, the Mixture-of-Diverse- Experts (MoDE) block is designed to learn the generalized parameters for all tasks, and gating re-parameterization (GatRep) is performed to generate the specialized param- eters for each task, by which RepMode can maintain a com- pact practical topology exactly like a plain network, and meanwhile achieves a powerful theoretical topology. Com- prehensive experiments show that RepMode can achieve state-of-the-art overall performance in SSP .	1. Introduction Recent years have witnessed great progress in biological research at the subcellular level [ 1,2,7,21,23,59,61], which plays a pivotal role in deeply studying cell functions and behaviors. To address the difficulty of observing subcellu- *Corresponding Author: [email protected] • • • (b)Partial Labeling (c)Multi-Scale• • • Transmitted-Light Image Fluorescent ImageNeural NetworkMicrotubule Golgi DNA Cell Membrane Mitochondrion Nucleolus Nuclear Envelope (a) × × × ×• • • × × Figure 1. (a) Illustration of subcellular structure prediction (SSP), which aims to predict the 3D fluorescent images of multiple sub- cellular structures from a 3D transmitted-light image. This task faces two challenges, i.e. (b) partial labeling and (c) multi-scale. lar structures, fluorescence staining was invented and has become a mainstay technology for revealing the locations and morphology of subcellular structures [ 26]. Specifically, biologists use the antibodies coupled to different fluores- cent dyes to ªstainº cells, after which the subcellular struc- tures of interest can be visualized by capturing distinct flu- orescent signals [ 64]. Unfortunately, fluorescence staining is expensive and time-consuming due to the need for ad- vanced instrumentation and material preparation [ 29]. Be- sides, phototoxicity during fluorescent imaging is detrimen- tal to living cells [ 28]. In this paper, we model fluorescence This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 3312 staining as a deep learning task, termed subcellular struc- ture prediction (SSP), which aims to directly predict the 3D fluorescent images of multiple subcellular structures from a 3D transmitted-light image (see Fig. 1(a)). The adoption of SSP can significantly reduce the expenditure on subcellular research and free biologists from this demanding workflow. Such an under-explored and challenging bioimage prob- lem deserves the attention of the computer vision commu- nity due to its high potential in biology. Specifically, SSP is a dense regression task where the fluorescent intensities of multiple subcellular structures need to be predicted for each transmitted-light voxel. However, due to the limita- tions of current biotechnology, each image can only obtain partial labels . For instance, some images may only have the annotations of nucleoli, and others may only have the anno- tations of microtubules (see Fig. 1(b)). Moreover, different subcellular structures would be presented at multiple scales under the microscope, which also needs to be taken into ac- count. For example, the mitochondrion is a small structure inside a cell, while obviously the cell membrane is a larger one since it surrounds a cell (see Fig. 1(c)). Generally, there are two mainstream solutions: 1) Multi- Net[5,32,33,47]: divide SSP into several individual predic- tion tasks and employs multiple networks; 2) Multi-Head [6,9,46]: design a partially-shared network composed of a shared feature extractor and multiple task-specific heads (see Fig. 2(a)). However, these traditional approaches or- ganize network parameters in an inefficient andinflexible manner, which leads to two major issues. First, they fail to make full use of partially labeled data in SSP, resulting in label-inefficiency . In Multi-Net, only the images contain- ing corresponding labels would be selected as the training set for each network and thus the other images are wasted, leading to an unsatisfactory generalization ability. As for Multi-Head, although all images are adopted for training, only partial heads are updated when a partially labeled im- age is input and the other heads do not get involved in train- ing. Second, to deal with the multi-scale nature of SSP, they require exhausting pre-design of the network architecture, and the resultant one may not be suitable for all subcellular structures, which leads to scale-inflexibility . In response to the above issues, herein we propose Re- parameterizing Mixture-of-Diverse-Experts (RepMode ), an all-shared network that can dynamically organize its pa- rameters with task-aware priors to perform specified single- label prediction tasks of SSP (see Fig. 2(b)). Specifically, RepMode is mainly constructed of the proposed Mixture-of- Diverse-Experts (MoDE )blocks . The MoDE block contains the expert pairs of various receptive fields, where these task- agnostic experts with diverse configurations are designed to learn the generalized parameters for all tasks . Moreover, gating re-parameterization (GatRep ) is proposed to conduct thetask-specific combinations of experts to achieve efficientexpert utilization, which aims to generate the specialized parameters for each task . With such a parameter organiz- ing manner (see Fig. 2(c)), RepMode can maintain a practi- cal topology exactly like a plain network, and meanwhile achieves a theoretical topology with a better representa- tional capacity. Compared to the above solutions, RepMode can fully learn from all training data, since the experts are shared with all tasks and thus participate in the training of each partially labeled image. Besides, RepMode can adap- tively learn the preference of each task for the experts with different receptive fields, thus no manual intervention is re- quired to handle the multi-scale issue. Moreover, by fine- tuning few newly-introduced parameters, RepMode can be easily extended to an unseen task without any degradation of the performance on the previous tasks. Our main contri- butions are summarized as follows: • We propose a stronger baseline for SSP, named Rep- Mode, which can switch different ªmodesº to predict multiple subcellular structures and also shows its po- tential in task-incremental learning. • The MoDE block is designed to enrich the generalized parameters and GatRep is adopted to yield the special- ized parameters, by which RepMode achieves dynamic parameter organizing in a task-specific manner. • Comprehensive experiments show that RepMode can achieve state-of-the-art (SOTA) performance in SSP. Moreover, detailed ablation studies and further analy- sis verify the effectiveness of RepMode.
Zhu_VDN-NeRF_Resolving_Shape-Radiance_Ambiguity_via_View-Dependence_Normalization_CVPR_2023	Abstract We propose VDN-NeRF , a method to train neural ra- diance fields (NeRFs) for better geometry under non- Lambertian surface and dynamic lighting conditions that cause significant variation in the radiance of a point when viewed from different angles. Instead of explicitly model- ing the underlying factors that result in the view-dependent phenomenon, which could be complex yet not inclusive, we develop a simple and effective technique that normalizes the view-dependence by distilling invariant information al- ready encoded in the learned NeRFs. We then jointly train NeRFs for view synthesis with view-dependence normal- ization to attain quality geometry. Our experiments show that even though shape-radiance ambiguity is inevitable, the proposed normalization can minimize its effect on geom- etry, which essentially aligns the optimal capacity needed for explaining view-dependent variations. Our method ap- plies to various baselines and significantly improves geom- etry without changing the volume rendering pipeline, even if the data is captured under a moving light source. Code is available at: https://github.com/BoifZ/VDN-NeRF.	1. Introduction Reconstructing the geometry and appearance of a 3D scene from a set of 2D images is one of the fundamen- tal tasks in computer vision and graphics. Recently, based on volume rendering, neural radiance fields (NeRFs) have shown great potential in capturing detailed appearance of the scene, evidenced by their high-quality view synthesis. However, the reconstructed geometry is still far from sat- isfying. Since geometry is critical for physical interaction leveraging such scene representations, many recent works have started to improve the geometric reconstruction within the context of volume rendering, e.g., with surface render- ing or external signals that are more informative of the ge- *Equal contributions. †Corresponding authors ([email protected], [email protected]). ‡The State Key Lab of CAD&CG. §The department of EEE and the Institute of Data Science. Input images Rendered image GeometryNeuS OurS Ground TruthFigure 1. We aim for quality geometry with NeRF reconstruction under inconsistency when viewing the same 3D point, for exam- ple, with images captured under a dynamic light field (top row), shown by the light spots on the truck cast from a torch moving with the camera. The middle two rows compare the reconstructed geometry from NeuS [41] (first column) and our method (second column). As observed, our method produces more details and bet- ter estimates the truck’s structure. The last row shows a novel view rendering from both methods. Even though our method normal- izes the view-dependence, it does not lose details on the synthe- sized images thanks to the regularity induced by better geometry. ometry, like depth scans. Nevertheless, most improvements do not explicitly study the shape-radiance ambiguity that could induce degener- ated geometric reconstructions. This ambiguity persists as long as some capacity is needed to account for the direc- tional variations in the radiance, which is further amplified This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 35 if the ambient light field changes according to the observer’s viewpoint. For example, the NeRF’s Multi-Layer Percep- trons (MLPs) takes in a 3D location and a 2D direction vec- tor and outputs the observed color of this point from this specific viewing angle. If the MLP has sufficient capacity to model the directional phenomenon, a perfect photomet- ric reconstruction could be achieved even if the learned ge- ometry is entirely wrong. In other words, wrong geometry incurs more directional variations, which the MLP’s view- dependent branch can still encode, thus, rendering the pho- tometric reconstruction loss incapable of constraining the solutions. On the other hand, directional capacity is needed to pre- vent the (view-dependent) photometric loss from distorting the geometry (confirmed through our experiments). With quality geometry as the central goal, we are now facing a tradeoff between the capacity of the view-dependent radi- ance function and the shape-radiance ambiguity. Namely, the MLP’s capacity has to be increased to explain the direc- tional variations, but if the capacity gets too large, shape- radiance ambiguity will permit degenerated geometric re- constructions. One can thus tune the capacity to achieve the best geometry for each scene, but it is time-consuming and infeasible since different scenes come with different levels of view dependence. Instead of tuning the directional capacity, we adjust the view-dependence. More explicitly, we propose to normalize the directional variations by encoding invariant features dis- tilled from the same scene representation. By doing so, the view-dependence of different scenes is aligned to the same level so that a single optimal capacity can ensure good view synthesis and prevent shape-radiance ambiguity from de- grading the geometric estimation. Given the self-distillation property, the proposed view-dependence normalization can be jointly trained with the commonly used photometric re- construction loss. It can also be easily plugged into any method relying on volume rendering for geometric scene reconstruction. We demonstrate the effectiveness of the proposed view- dependence normalization on several datasets. Especially we verify that the level of view-dependence affects the min- imality of the shape-radiance ambiguity with tunable direc- tional capacity. And we show that our method effectively aligns the optimal capacity for each scene. To evaluate how our method works under a dynamically changing light field, we propose a new benchmark where a light source is mov- ing conditioned on the camera pose. We validate that the geometry obtained from our method degrades gracefully as the significance of the induced directional variations in- creases. In summary, we make the following contributions: • We perform a detailed study of the coupling between the capacity of the view-dependent radiance function and the shape-radiance ambiguity in the context of vol-ume rendering. • We propose a view-dependence normalization method that effectively aligns the optimality of the direc- tional function capacity under shape-radiance ambigu- ity for each scene and achieves state-of-the-art geome- try compared to various baselines. • We quantitatively verify the robustness of our method when the light field changes to further increase the di- rectional variation, ensuring the applicability to sce- narios with unfavored lighting conditions.
Zhao_Improved_Distribution_Matching_for_Dataset_Condensation_CVPR_2023	Abstract Dataset Condensation aims to condense a large dataset into a smaller one while maintaining its ability to train a well-performing model, thus reducing the storage cost and training effort in deep learning applications. However, con- ventional dataset condensation methods are optimization- oriented and condense the dataset by performing gradient or parameter matching during model optimization, which is computationally intensive even on small datasets and mod- els. In this paper, we propose a novel dataset condensation method based on distribution matching, which is more ef- ficient and promising. Specifically, we identify two impor- tant shortcomings of naive distribution matching ( i.e., im- balanced feature numbers and unvalidated embeddings for distance computation) and address them with three novel techniques ( i.e., partitioning and expansion augmentation, efficient and enriched model sampling, and class-aware dis- tribution regularization). Our simple yet effective method outperforms most previous optimization-oriented methods with much fewer computational resources, thereby scal- ing data condensation to larger datasets and models. Ex- tensive experiments demonstrate the effectiveness of our method. Codes are available at https://github. com/uitrbn/IDM	1. Introduction Deep learning [23, 25, 57] is notoriously data-hungry, which poses challenges for both its training and data stor- age. To improve data storage efficiency, Dataset Conden- sation (DC) [47, 56] aims to condense large datasets into smaller ones while retaining their validity for model train- ing. Unlike traditional coreset selection methods [38, 44, 49,50], such dataset condensation is often achieved through image synthesis and yields better performance. If prop- erly condensed, the resulting datasets not only consume less storage space, but can also benefit various downstream tasks *Corresponding authors are Guanbin Li and Yizhou Yu. Optimization - Oriented Methods Distribution Matching Methods : Model Parameter at Iteration t Forward Backward & : Synthetic and real samples : Loss of Target Task : Gradient : Model OutputFigure 1. Illustration of optimization-oriented methods and distri- bution matching methods. L: classification loss; g: gradient; O: output of models; Lmatching : the matching loss for condensation. such as network architecture search and continual learning by reducing their computational costs. The pioneering DC approach [56, 56] guarantees the va- lidity of a condensed dataset by imposing a strong assump- tion that a model trained on it should be identical to that trained on the real dataset. However, naive matching be- tween these converged models can be too challenging due to their large parameter space and long optimization towards convergence. To this end, they impose an even stronger as- sumption that the two models should share an identical or similar optimization path, which can be achieved by match- ing either their gradients [53,56] or their intermediate model parameters [8] during training. We therefore refer to them asoptimization-oriented methods. However, despite their success, the unique characteris- tics of optimization-oriented DC methods imply that they inevitably suffer from high computational costs and thus scale poorly to large datasets and models. Specifically, they all involve the optimization of models (either randomly- initialized [53, 56] or initialized with parameters of pre- trained models [8]) against the condensed dataset, and thus rely on a nested loop that optimizes the condensed dataset and model parameters in turn. Note that the use of pre- trained models require additional computation and storage space [8]. As a result, existing optimization-oriented DC methods are only applicable to “toy” networks ( e.g., three- layer convolutional networks) and small datasets ( e.g., CI- FAR10, CIFAR100). Whether they can be scaled to real- world scenarios is still an open question. This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 7856 To scale DC to large models and datasets, distribution matching (DM) [54] proposes to match the output feature distributions of the real and condensed datasets extracted by randomly-initialized models. This stems from the fact that the validity of a condensed dataset can also be guar- anteed if it produces the same feature distribution as the real dataset. Since DM does not involve the optimization of models against the condensed dataset, it avoids the expen- sive nested loops in optimization-oriented methods, and is thus highly efficient and scalable. However, despite being promising, experimental results show that its performance still lags behind that of the state-of-the-art optimization- oriented methods. In this paper, we perform an in-depth analysis on DM’s unsatisfactory performance and propose a set of remedies that can significantly improve it, namely improved distribu- tion matching (IDM). Specifically, we analyzed the output feature distributions of DM and observed that although their means match, the features of the condensed dataset scatter around, causing severe class misalignment problems, which accounts for its impaired performance. We ascribe such scattered features to two shortcomings of DM as follows: i) DM suffers from the imbalanced number of features. In- tuitively, DM uses the features of small condensed datasets to match those of large real datasets, which is inherently intractable. Addressing this shortcoming, we propose Par- titioning and Expansion augmentation , which augments the condensed dataset by evenly splitting each image into l×l parts and expanding each part to the size of the original im- age, resulting in l2features per image and a better match to the features of the real dataset. ii) Randomly initialized models are not valid embedding functions for the Maximum Mean Discrepancy (MMD) [18] estimation used in DM. Specifically, DM justifies the valid- ity of randomly-initialized models by their intrinsic clas- sification power observed in tasks such as deep cluster- ing [3,5,6,36]. However, we believe that this does not apply to DM as randomly-initialized models do not satisfy the re- quirement of embedding functions used in MMD and makes it an invalid measure of distribution distance. Since it is too challenging to design neural network based embedding functions that are valid for MMD, we propose two simple yet effective remedies: 1) Efficient and enriched model sam- pling . We enrich the embedding functions in MMD with semi-trained models as additional feature extractors, and develop a memory-efficient model queue to facilitate their sampling. 2) Class-aware distribution regularization . We explicitly regularize the feature distributions of condensed datasets to further alleviate class misalignment. These three novel techniques together help to extract better feature distributions for DM. Our contributions include: • We deeply analyze the shortcomings of the Distributed Matching [54] algorithm and reveal that the root of itsimpaired performance lies in the problem of class mis- alignment. • We propose improved distribution matching (IDM), consisting of three novel techniques that address the shortcomings of DM and help to learn better feature distributions. • Experimental results show that our IDM achieves sig- nificant improvement over DM and surpasses the per- formance of most optimization-oriented methods. • We show that our IDM method is highly efficient and scalable, and can be applied to large datasets such as ImageNet Subset [10, 43].
Zuo_Natural_Language-Assisted_Sign_Language_Recognition_CVPR_2023	Abstract Sign languages are visual languages which convey in- formation by signers’ handshape, facial expression, body movement, and so forth. Due to the inherent restric- tion of combinations of these visual ingredients, there ex- ist a significant number of visually indistinguishable signs (VISigns) in sign languages, which limits the recognition capacity of vision neural networks. To mitigate the problem, we propose the Natural Language-Assisted Sign Language Recognition (NLA-SLR) framework, which exploits seman- tic information contained in glosses (sign labels). First, for VISigns with similar semantic meanings, we propose language-aware label smoothing by generating soft labels for each training sign whose smoothing weights are com- puted from the normalized semantic similarities among the glosses to ease training. Second, for VISigns with distinct semantic meanings, we present an inter-modality mixup technique which blends vision and gloss features to further maximize the separability of different signs under the super- vision of blended labels. Besides, we also introduce a novel backbone, video-keypoint network, which not only models both RGB videos and human body keypoints but also de- rives knowledge from sign videos of different temporal re- ceptive fields. Empirically, our method achieves state-of- the-art performance on three widely-adopted benchmarks: MSASL, WLASL, and NMFs-CSL. Codes are available at https://github.com/FangyunWei/SLRT.	1. Introduction Sign languages are the primary languages for communi- cation among deaf communities. On the one hand, sign lan- guages have their own linguistic properties as most natural languages [1,52,64]. On the other hand, sign languages are visual languages that convey information by the movements of the hands, body, head, mouth, and eyes, making them completely separate and distinct from natural languages [6,69,71]. This work dedicates to sign language recognition (SLR), which requires models to classify the isolated signs †Corresponding author. Gloss: “C old” Gloss: “Winter”(a) VISigns may have similar semantic meanings. Gloss: “Afternoon”Gloss: “Table” (b) VISigns may have distinct semantic meanings. Figure 1. Vision neural networks are demonstrated to be less ef- fective to recognize visually indistinguishable signs (VISigns) [2, 26, 34]. We observe that VISigns may have similar or distinct se- mantic meanings, inspiring us to leverage this characteristic to fa- cilitate sign language recognition as illustrated in Figure 2. from videos into a set of glosses1. Despite its fundamen- tal capacity of recognizing signs, SLR has a broad range of applications including sign spotting [36, 42, 58], sign video retrieval [8, 11], sign language translation [6, 35, 54], and continuous sign language recognition [1, 6]. Since the lexical items of sign languages are defined by the handshape, facial expression, and movement, the combinations of these visual ingredients are restricted in- herently, yielding plenty of visually indistinguishable signs termed VISigns. VISigns are those signs with similar hand- shape and motion but varied semantic meanings. We show two examples (“Cold” vs. “Winter” and “Table” vs. “Af- ternoon”) in Figure 1. Unfortunately, it has been demon- strated that vision neural networks are less effective at 1Gloss is a unique label for a single sign. Each gloss is identified by a word which is associated with the sign’s semantic meaning. This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 14890 V ocab: {“Above”, …, “Zero”} “Twenty”… SimilaritiesSoft Label of “Twenty”Gloss Feature(a) Language-aware label smoothing. “Twenty”V ocab: {“Above”, …, “Zero”} … Vision FeatureAddition Operator “Twenty” or “Above”?“Twenty” or “Zero”?… Gloss Features (b) Inter-modality mixup. Figure 2. We incorporate natural language modeling into sign language recognition to promote recognition capacity. (a) Language-aware label smoothing generates a soft label for each training video, whose smoothing weights are the normalized semantic similarities of the ground truth gloss and the remaining glosses within the sign language vocabulary. (b) Inter-modality mixup yields the blended features (denoted by orange rectangles) with the corresponding mixed labels to maximize the separability of signs in a latent space. accurately recognizing VISigns [2, 26, 34]. Due to the intrinsic connections between sign languages and natural languages, the glosses, i.e., labels of signs, are seman- tically meaningful in contrast to the one-hot labels used in traditional classification tasks [27, 51]. Thus, although it is challenging to classify the VISigns from the vision perspective, their glosses provide serviceable semantics, which is, however, less taken into consideration in previ- ous works [18–20,23,24,26,34,36]. Our work is built upon the following two findings. Finding-1: VISigns may have similar semantic meanings (Figure 1a). Due to the observation that VISigns may have higher visual similarities, assigning hard labels to them may hinder the training since it is challenging for vision neu- ral networks to distinguish each VISign apart. A straight- forward way to ease the training is to replace the hard la- bels with soft ones as in well-established label smooth- ing [15, 56]. However, how to generate proper soft labels is non-trivial. The vanilla label smoothing [15, 56] assigns equal smoothing weights to all negative terms, which ig- nores the semantic information contained in labels. In light of the finding-1 that VISigns may have similar semantic meanings and the intrinsic connections between sign lan- guages and natural languages, we consider the semantic similarities among the glosses when generating soft labels. Concretely, for each training video, we adopt an off-the- shelf word representation framework, i.e., fastText [39], to pre-compute the semantic similarities of its gloss and the re- maining glosses within the sign language vocabulary. Then we can properly generate a soft label for each training sam- ple whose smoothing weights are the normalized semantic similarities. In this way, negative terms with similar seman- tic meanings to the ground truth gloss are assigned higher values in the soft label. As shown in Figure 2a, we term this process as language-aware label smoothing, which injects prior knowledge into the training. Finding-2: VISigns may have distinct semantic mean- ings (Figure 1b). Although the VISigns are challenging to be classified from the vision perspective, the semantic meanings of their glosses may be distinguishable according tofinding-2 . This inspires us to combine the vision fea-tures and gloss features to drive the model towards max- imizing signs’ separability in a latent space. Specifically, given a sign video, we first leverage our proposed backbone to encode its vision feature and the well-established fast- Text [39] to extract the feature of each gloss within the sign language vocabulary. Then we independently integrate the vision feature and each gloss feature to produce a blended representation, which is further fed into a classifier to ap- proximate its mixed label. We refer to this procedure as inter-modality mixup as shown in Figure 2b. We empir- ically find that our inter-modality mixup significantly en- hances the model’s discriminative power. Our contributions can be summarized as follows: • We are the first to incorporate natural language mod- eling into sign language recognition based on the dis- covery of VISigns. Language-aware label smoothing and inter-modality mixup are proposed to take full ad- vantage of the linguistic properties of VISigns and se- mantic information contained in glosses. • We take into account the unique characteristic of sign languages and present a novel backbone named video- keypoint network (VKNet), which not only models both RGB videos and human keypoints, but also de- rives knowledge from sign videos of various temporal receptive fields. • Our method, termed natural language-assisted sign language recognition (NLA-SLR), achieves state-of- the-art performance on the widely-used SLR datasets including MSASL [26], WLASL [34], and NMFs- CSL [20].
Zhou_ZegCLIP_Towards_Adapting_CLIP_for_Zero-Shot_Semantic_Segmentation_CVPR_2023	Abstract Recently, CLIP has been applied to pixel-level zero-shot learning tasks via a two-stage scheme. The general idea is to ﬁrst generate class-agnostic region proposals and thenfeed the cropped proposal regions to CLIP to utilize itsimage-level zero-shot classiﬁcation capability. While ef-fective, such a scheme requires two image encoders, onefor proposal generation and one for CLIP , leading to acomplicated pipeline and high computational cost. In thiswork, we pursue a simpler-and-efﬁcient one-stage solu- tion that directly extends CLIP’s zero-shot prediction ca-pability from image to pixel level. Our investigation starts with a straightforward extension as our baseline that gen-erates semantic masks by comparing the similarity betweentext and patch embeddings extracted from CLIP . However , such a paradigm could heavily overﬁt the seen classes andfail to generalize to unseen classes. To handle this is-sue, we propose three simple-but-effective designs and ﬁg- ure out that they can signiﬁcantly retain the inherent zero-shot capacity of CLIP and improve pixel-level generaliza- tion ability. Incorporating those modiﬁcations leads toan efﬁcient zero-shot semantic segmentation system calledZegCLIP . Through extensive experiments on three public benchmarks, ZegCLIP demonstrates superior performance,outperforming the state-of-the-art methods by a large mar-gin under both “inductive” and “transductive” zero-shotsettings. In addition, compared with the two-stage method,our one-stage ZegCLIP achieves a speedup of about 5 times faster during inference. We release the code at https: //github.com/ZiqinZhou66/ZegCLIP.git .	1. Introduction Semantic segmentation is one of the fundamental tasks in the computer vision ﬁeld, which aims to predict thecategory of each pixel of an image [ 7,15,31,41]. Ex- tensive works have been proposed [ 8,26,30], e.g., Fully Convolutional Networks [ 29], U-net [ 38], DeepLab fam- ily [ 4–6] and more recently Vision Transformer based meth- ods [ 16,53,58]. *Corresponding author71.176.5 75.989.9 91.9 16.3 13.828.340.477.8 020406080100 mIoU(%)Seen Unseen Figure 1. Quantitative improvements achieved by our proposed designs on VOC dataset. (1)(2) represents our one-stage Base- line model of different versions (Fix or Fine-Tune CLIP image en- coder), while (3)-(5) shows the effectiveness of applying our pro-posed designs, i.e., Deep Prompt Tuning (DPT) ,Non-mutually Exclusive Loss (NEL) ,Relationship Descriptor (RD) , on base- line model step by step. We highlight that our designs can dramat-ically increase the segmentation performance on unseen classes. However, the success of the deep semantic segmentation models heavily relies on the availability of a large amountof annotated training images, which involves a substantialamount of labor. This gives rise to a surging interest in low-supervision-based semantic segmentation approaches, in- cluding semi-supervised [ 7], weakly-supervised [ 48], few- shot [ 46], and zero-shot semantic segmentation [ 3,34,44]. Among them, zero-shot semantic segmentation is partic- ularly challenging and attractive since it is required to di-rectly produce the semantic segmentation results based on the semantic description of a given class. Recently, the pre-trained vision-language model CLIP [ 36] has been adopted into various dense prediction tasks, such as referring seg-mentation [ 43], semantic segmentation [ 33], and detection [14]. It also offers a new paradigm and has made a break- through for zero-shot semantic segmentation. Initially builtfor matching text and images, CLIP has demonstrated a re- markable capability for image-level zero-shot classiﬁcation. However, zsseg [ 49] and Zegformer [ 12] follow the com- mon strategy that needs two-stage processing that ﬁrst gen- erates region proposals and then feeds the cropped regions to CLIP for zero-shot classiﬁcation. Such a strategy re-quires two image encoding processes, one for generatingproposals and one for encoding each proposal via CLIP . This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 11175 Table 1. Differences between our approach and related zero-shot semantic segmentation methods based on CLIP . MethodsNeed an extra CLIP as an image Can do image encoder? -level classiﬁer? inductive? zsseg [ 49] /enc-33/enc-33 /enc-33 ZegFormer [ 12] /enc-33/enc-33 /enc-33 MaskCLIP+ [ 56] /enc-33/enc-37 /enc-37 ZegCLIP (Ours) /enc-37/enc-37 /enc-33 This design creates additional computational overhead and cannot leverage the knowledge of the CLIP encoder at theproposal generation stage. Besides, MaskCLIP+ [ 56] uti- lizes CLIP to generate pseudo labels of novel classes forself-training but will be invalid if the unseen class namesin inference are unknown in the training stage (“inductive”zero-shot setting). This paper pursues simplifying the pipeline by directly extending the zero-shot capability of CLIP from image- level to pixel-level. The basic idea of our method is straight-forward: we use a lightweight decoder to match the textprompts against the local embeddings extracted from CLIP ,which could be achieved via the self-attention mechanism in a transformer-based structure. We train the vanilla decoderand ﬁx or ﬁne-tune the CLIP image encoder on a datasetcontaining pixel-level annotations from a limited number of classes, expecting the text-patch matching capability cangeneralize to unseen classes. Unfortunately, this basic ver-sion tends to overﬁt the training set: while the segmentationresults for seen classes generally improve, the model failsto produce reasonable segments on unseen classes. Sur-prisingly, we discover such an overﬁtting issue can be dra-matically alleviated by incorporating three modiﬁed designchoices and report the quantitative improvements in Fig. 1. The following highlights our key discoveries: Design 1: Using Deep Prompt Tuning (DPT) instead of ﬁne-tuning or ﬁxing for the CLIP image encode. We ﬁnd that ﬁne-tuning could lead to overﬁtting to seen classeswhile prompt tuning prefers to retain the inherent zero-shotcapacity of CLIP . Design 2: Applying Non-mutually Exclusive Loss (NEL) function when performing pixel-level classiﬁcation but gen-erating the posterior probability of one class independent ofthe logits of other classes. Design 3: Most importantly and our major innovation — introducing a Relationship Descriptor (RD) to incorporate the image-level prior into text embedding before matchingtext-patch embeddings from CLIP can signiﬁcantly preventthe model from overﬁtting to the seen classes. By incorporating those three designs into our one-stage baseline, we create a simple-but-effective zero-shot seman-tic segmentation model named ZegCLIP . Tab. 1summa- rizes the differences between our proposed method and ex-isting approaches based on CLIP . More details can be foundin Appendix. We conduct extensive experiments on threepublic datasets and show that our method outperforms the state-of-the-art methods by a large margin in both the “in-ductive” and “transductive” settings.
Zhao_High-Frequency_Stereo_Matching_Network_CVPR_2023	Abstract In the field of binocular stereo matching, remarkable progress has been made by iterative methods like RAFT- Stereo and CREStereo. However, most of these methods lose information during the iterative process, making it difficult to generate more detailed difference maps that take full advantage of high-frequency information. We propose the Decouple module to alleviate the problem of data coupling and allow features containing subtle details to transfer across the iterations which proves to alleviate the problem significantly in the ablations. To further capture high-frequency details, we propose a Normalization Refine- ment module that unifies the disparities as a proportion of the disparities over the width of the image, which address the problem of module failure in cross-domain scenarios. Further, with the above improvements, the ResNet-like feature extractor that has not been changed for years becomes a bottleneck. Towards this end, we proposed a multi-scale and multi-stage feature extractor that intro- duces the channel-wise self-attention mechanism which greatly addresses this bottleneck. Our method (DLNR) ranks1ston the Middlebury leaderboard, significantly outperforming the next best method by 13.04%. Our method also achieves SOTA performance on the KITTI- 2015 benchmark for D1-fg. Code and demos are available at:https://github.com/David-Zhao-1997/ High-frequency-Stereo-Matching-Network . †These authors contributed equally. ⋆Corresponding author. Email: [email protected]. ⋆⋆Second Corresponding author. Email: [email protected]	1. Introduction Figure 1. Motivation. We aim to address the problem of blurry edges, thin object missing, and textureless region mismatch. Stereo depth estimation is becoming the infrastructure for 3D applications. Accurate depth perception is vital for autonomous driving, drones navigation, robotics and other related fields. The main point of the task is to estimate a pixel-wise displacement map also known as disparity that can be used to determine the depth of the pixels in the scene. Traditional stereo matching algorithms [ 7,11,12] are mainly divided into two types: global methods [ 6,16,17,26] and local methods [ 1,13]. Both methods solve the optimiza- tion problem by minimizing the objective function contain- ing the data and smoothing terms, while the former takes into account the global information, the latter simply takes into account the local information, hence both have their own benefits in terms of accuracy and speed when solv- ing the optimization problem. Traditional methods have ex- cellent generalization performance and robustness in differ- ent scenarios, but perform poorly on details such as weak This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 1327 textures and repetitive texture regions. With the develop- ment of convolutional neural networks, learning-based ap- proaches [ 20,28,37] have lately demonstrated promising result in tackling the matching problem of challenging re- gions. Take advantages of the strong regularization perfor- mance of the 3D convolution and 4D cost volume, meth- ods [ 2,10,15,45] using 3D convolution performs well. While their practical applicability is limited by the high computational cost. Subsequent methods [ 37,43] attempt to use multiple adaptive aggregated and guided aggregated 2D convolutions instead of 3D convolution, reducing computa- tional cost and achieving better performance. The recent ap- pearance of RAFT-Stereo [ 20] has given rise to a fresh con- cept for the research of stereo matching. Derived from the optical estimation method RAFT [ 29], RAFT-Stereo uses the iterative refinement method for a coarse-to-fine pipeline. It first calculates the correlation between all pixel pairs to construct a 3D correlation pyramid. Then an update oper- ator with a convolutional GRU as the core unit is used to retrieves features from the correlation pyramid and updates the disparity map [ 20]. Despite great progress has been made in learning-based approaches, two major problems remain. (1) Most current approaches fall short when it comes to the finer features of the estimated disparity map. Especially for the edge perfor- mance of the objects. In bokeh and rendering applications, the edge performance of the disparity map is critical to the final result. For example, technologies that require pixel- level rendering, such as VR and AR, have high requirements for fitting between the scene model and the image mapping, which means we need a tight fit between the edges in the disparity map and the original RGB image. (2) The mis- match of textureless regions and the missing of thin objects are also important factors that significantly deteriorate the disparity map. For example, the mismatch of weak texture walls and the missing of thin electrical wires are fatal flaws for obstacle avoidance applications. To alleviate these problems, we propose DLNR (Stereo Matching Network with Decouple LSTM and Normaliza- tion Refinement), a new end-to-end data-driven method for stereo matching. We introduced several improvements based on the itera- tive model: Most of the current iterative methods usually apply the original GRU structure as their iterative cell. While the problem is that in the original GRU structure, the informa- tion used to generate the update matrix of the disparity map is coupled with the value of the hidden state transfer be- tween iterations, making it hard to keep subtle details in the hidden state. Therefore, we designed the Decouple LSTM module to decouple the hidden state from the update ma- trix of the disparity map. Experiments and visualizations proved that the module retains more subtle details in thehidden states. Decouple LSTM keeps more high-frequency informa- tion in the iterative stage through data decoupling, how- ever, in order to balance performance and computational speed, the resolution of the iterative stage is only 1/4 of the original resolution at most. To produce disparity maps with sharp edges and subtle details, a subsequent refine- ment module is still needed. In our refinement module, we aim to sufficiently exploit the information from the upsam- pled disparity maps, the original left and right images con- taining high-frequency information to enhance edges and details. However, due to the large differences in disparity ranges between different images and different datasets, the Refinement module often has poor generalization perfor- mance when encountering images with different disparity ranges. In particular, when performing finetune, the mod- ule may even fail when encountering disparity ranges that differ greatly. To address this problem, we propose the Dis- parity Normalization strategy. Experiments and visualiza- tions proved that the module improves performance as well as alleviates the problem of domain difference. After the above two improvements, we found that the feature extractor became the bottleneck of the performance. In the field of stereo matching, feature extraction has not been improved significantly for years, most learning-based methods still use ResNet-like feature extractors which fall short when providing information for well-designed post- stage structures. To alleviate the problem, we propose the Channel-Attention Transformer feature extractor aims to capture long-range pixel dependencies and preserve high- frequency information.
Zhu_I2-SDF_Intrinsic_Indoor_Scene_Reconstruction_and_Editing_via_Raytracing_in_CVPR_2023	Abstract In this work, we present I2-SDF , a new method for in- trinsic indoor scene reconstruction and editing using dif- ferentiable Monte Carlo raytracing on neural signed dis- tance fields (SDFs). Our holistic neural SDF-based frame- work jointly recovers the underlying shapes, incident ra- diance and materials from multi-view images. We intro- duce a novel bubble loss for fine-grained small objects and error-guided adaptive sampling scheme to largely improve the reconstruction quality on large-scale indoor scenes. Further, we propose to decompose the neural radiance field into spatially-varying material of the scene as a neu- ral field through surface-based, differentiable Monte Carlo raytracing and emitter semantic segmentations, which en- ables physically based and photorealistic scene relighting and editing applications. Through a number of qualita- tive and quantitative experiments, we demonstrate the su- perior quality of our method on indoor scene reconstruc- tion, novel view synthesis, and scene editing compared to state-of-the-art baselines. Our project page is at https: //jingsenzhu.github.io/i2-sdf .	1. Introduction Reconstructing 3D scenes from multi-view images is a fundamental task in computer graphics and vision. Neu- ral Radiance Field (NeRF) [16] and its follow-up research leverage multi-layer perceptions (MLPs) as implicit func- tions, taking as input the positional and directional coordi- nates, to approximate the underlying geometry and appear- ance of a 3D scene. Such methods have shown compelling and high-fidelity results in novel view synthesis. However, we argue that novel view synthesis itself is insufficient for scene editing such as inserting virtual objects, relighting *Corresponding author.RENDERING SCENE EDITING Figure 1. I2-SDF. Left: State-of-the-art neural implicit surface representation method [42] fails in reconstructing small objects inside an indoor scene ( e.g. lamps and chandeliers), which is re- solved by our bubbling method. Middle and Right: Our intrinsic decomposition and raytracing method enable photo-realistic scene editing and relighting applications. and editing surface materials with global illumination. On the other hand, inverse rendering or intrinsic de- composition , which reconstructs and decomposes the scene into shape, shading and surface reflectance from single or multiple images, enables photorealistic scene editing pos- sibilities. It is a long-term challenge especially for large- scale indoor scenes because they typically exhibit complex geometry and spatially-varying global illumination appear- ance. As intrinsic decomposition is an extremely ill-posed task, a physically-based shading model will crucially af- fect the decomposition quality. Existing neural rendering methods [2, 18, 43, 46] rely on simple rendering algorithms (such as pre-filtered shading) for the decomposition and use a global lighting representation (e.g., spherical Gaussians). Although these methods have demonstrated the effective- ness on object-level inverse rendering, they are inapplicable to complex indoor scenes. Moreover, indoor scene images are usually captured from the inside out and most lighting information has already presented inside the room. As a re- sult, the reconstructed radiance field already provides suffi- cient lighting information without the need of active, exter- nal capture lighting setup. To tackle the above challenges, we propose I2-SDF , a new method to decompose a 3D scene into its underlying This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 12489 shape, material, and incident radiance components using implicit neural representations. We design a robust two- stage training scheme that first reconstructs a neural SDF with radiance field, and then conducts raytracing in the SDF to decompose the radiance field into material and emission fields. As complex indoor scenes typically contain many fine-grained, thin or small structures with high-frequency details that are difficult for an implicit SDF function to fit, we propose a novel bubble loss and an error-guided adap- tive sampling scheme that greatly improve the reconstruc- tion quality on small objects in the scene. As a result, our approach achieves higher reconstruction quality in both geometry and novel view synthesis, outperforming previ- ous state-of-the-art neural rendering methods in complex indoor scenes. Further, we present an efficient intrinsic de- composition method that decomposes the radiance field into spatially-varying material and emission fields using surface- based, differentiable Monte Carlo raytracing, enabling var- ious scene editing applications. In summary, our contributions include: • We introduce I2-SDF1, a holistic neural SDF-based framework for complex indoor scenes that jointly re- covers the underlying shape, radiance, and material fields from multi-view images. • We propose a novel bubble loss and error-guided adap- tive sampling strategy to effectively reconstruct fine- grained small objects inside the scene. • We are the first that introduce Monte Carlo raytracing technique in scene-level neural SDF to enable photo- realistic indoor scene relighting and editing. • We provide a high-quality synthetic indoor scene multi-view dataset, with ground truth camera pose and geometry annotations.
Zheng_NeuralPCI_Spatio-Temporal_Neural_Field_for_3D_Point_Cloud_Multi-Frame_Non-Linear_CVPR_2023	Abstract In recent years, there has been a significant increase in focus on the interpolation task of computer vision. Despite the tremendous advancement of video interpolation, point cloud interpolation remains insufficiently explored. Mean- while, the existence of numerous nonlinear large motions in real-world scenarios makes the point cloud interpolation task more challenging. In light of these issues, we present NeuralPCI : an end-to-end 4D spatio-temporal Neural field for 3D PointCloud Interpolation, which implicitly inte- grates multi-frame information to handle nonlinear large motions for both indoor and outdoor scenarios. Further- more, we construct a new multi-frame point cloud interpo- lation dataset called NL-Drive for large nonlinear motions in autonomous driving scenes to better demonstrate the su- periority of our method. Ultimately, NeuralPCI achieves state-of-the-art performance on both DHB (Dynamic Hu- man Bodies) and NL-Drive datasets. Beyond the interpola- tion task, our method can be naturally extended to point cloud extrapolation, morphing, and auto-labeling, which indicates its substantial potential in other domains. Codes are available at https://github.com/ispc-lab/NeuralPCI.	1. Introduction In the field of computer vision, sequential point clouds are frequently utilized in many applications, such as VR/AR techniques [11, 38, 49] and autonomous driving [4, 32, 46]. The relatively low frequency of LiDAR compared to other sensors, i.e., 10–20 Hz, impedes exploration for high tem- poral resolution point clouds [47]. Therefore, interpolation tasks for point cloud sequences, which have not been sub- stantially investigated, are receiving increasing attention. With the similar goal of obtaining a smooth sequence with high temporal resolution, we can draw inspiration from the video frame interpolation (VFI) task. Several VFI meth- ods [6,8,20,35,45,48] concentrate on nonlinear movements in the real world. They take multiple frames as input and ∗Equal contribution.†Corresponding author. Figure 1. Common cases of nonlinear motions in autonomous driving scenarios. Spatially uniform linear interpolation ( ) using the middle two frames of the point cloud differs significantly from the actual situation ( ), so it is necessary to take multiple point clouds into consideration for nonlinear interpolation. generate explicit multi-frame fusion results based on flow estimation [6, 8, 20, 45, 48] or transformer [35]. Nonethe- less, due to the unique structure of point clouds [31], it is non-trivial to extend VFI methods to the 3D domain. Some early works [18,19] rely on stereo images to gener- ate pseudo-LiDAR point cloud interpolation. For pure point cloud input, previous methods [22,47] take two consecutive frames as input and output the point cloud at a given inter- mediate moment. However, with limited two input frames, these approaches can only produce linear interpolation re- sults [22], or perform nonlinear compensation by fusing in- put frames in the feature dimension linearly [47], which is inherently a data-driven approach to learning the dataset- specified distribution of nonlinear motions rather than an actual nonlinear interpolation. Only when the frame rate of the input point cloud sequence is high enough or the object motion is small enough, can the two adjacent point clouds satisfy the linear motion assumption. Nonetheless, there are numerous nonlinear motions in real-world cases. For instance, as illustrated in Fig. 1, the result of linear interpo- lation between two adjacent point cloud frames has a large deviation from the actual situation. A point cloud sequence rather than just two point cloud frames allows us to view further into the past and future. Neighboring multiple point This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 909 clouds contain additional spatial-temporal cues, namely dif- ferent perspectives, complementary geometry, and implicit high-order motion information. Therefore, it is time to re- think the point cloud interpolation task with an expanded design space, which is an open challenge yet. Methods that explicitly fuse multiple point cloud frames generally just approximate the motion model over time, which actually simplifies real-world complex motion. The neural field provides a more elegant way to parameterize the continuous point cloud sequence implicitly. Inspired by NeRF [25] whose view synthesis of images is essentially an interpolation, we propose NeuralPCI, a neural field to exploit the spatial and temporal information of multi-frame point clouds. We build a 4D neural spatio-temporal field, which takes sequential 3D point clouds and the indepen- dent interpolation time as input, and predicts the in-between or future point cloud at the given time. Moreover, Neu- ralPCI is optimized on runtime in a self-supervised man- ner, without relying on costly ground truths, which makes it free from the out-of-the-distribution generalization prob- lem. Our method can be flexibly applied to segmentation auto-labeling and morphing. Besides, we newly construct a challenging multi-frame point cloud interpolation dataset called NL-Drive from public autonomous driving datasets. Finally, we achieve state-of-the-art performance on indoor DHB dataset and outdoor NL-Drive dataset. Our main contributions are summarized as follows: • We propose a novel multi-frame point cloud interpola- tion algorithm to deal with the nonlinear complex mo- tion in real-world indoor and outdoor scenarios. • We introduce a 4D spatio-temporal neural field to in- tegrate motion information implicitly over space and time to generate the in-between point cloud frames at the arbitrary given time. • A flexible unified framework to conduct both the inter- polation and extrapolation, facilitating several applica- tions as well.
Zhu_PMatch_Paired_Masked_Image_Modeling_for_Dense_Geometric_Matching_CVPR_2023	Abstract Dense geometric matching determines the dense pixel- wise correspondence between a source and support image corresponding to the same 3D structure. Prior works em- ploy an encoder of transformer blocks to correlate the two- frame features. However, existing monocular pretraining tasks, e.g., image classification, and masked image model- ing (MIM), can not pretrain the cross-frame module, yield- ing less optimal performance. To resolve this, we reformu- late the MIM from reconstructing a single masked image to reconstructing a pair of masked images, enabling the pre- training of transformer module. Additionally, we incorpo- rate a decoder into pretraining for improved upsampling results. Further, to be robust to the textureless area, we pro- pose a novel cross-frame global matching module (CFGM). Since the most textureless area is planar surfaces, we pro- pose a homography loss to further regularize its learning. Combined together, we achieve the State-of-The-Art (SoTA) performance on geometric matching. Codes and models are available at https://github.com/ShngJZ/PMatch.	1. Introduction When a 3D structure is viewed in both a source and a support image, for a pixel (or keypoint) in the source image, the task of geometric matching identifies its corresponding pixel in the support image. This task is a cornerstone for many downstream vision applications, e.g.homography es- timation [18], structure-from-motion [45], visual odometry estimation [21] and visual camera localization [7]. There exist both sparse and dense methods for geomet- ric matching. The sparse methods [16, 19, 32, 33, 40, 42, 48, 48, 56] only yield correspondence on sparse or semi- dense locations while the dense methods [20, 54, 55] es- timate pixel-wise correspondence. They primarily differ in that the sparse methods embed a keypoint detection or a global matching on discrete coordinates, which underly- ingly assumes a unique mapping between source and sup- port frames. Yet, the existence of textureless surfaces in- Figure 1. Most vision tasks start with a pretrained network. In geo- metric matching, the unique network components processing two- view features cannot benefit from the monocular pretraining task, e.g., image classification, and masked image modeling (MIM). As in the figure, this work enables the pretraining of a matching model via reformulating MIM from reconstructing a single masked im- age to reconstructing a pair of masked images. troduces multiple similar local patches, disabling keypoint detection or causing ambiguous matching results. Dense methods, though facing similar challenges at the coarse level, alleviate it with the additional fine-level local context and smoothness constraint. Until recently, the dense meth- ods demonstrate a comparable or better geometric matching performance over the sparse methods [20, 54, 55]. A relevant task to dense geometric matching is the opti- cal flow estimation [50]. Both tasks estimate dense corre- spondences, whereas the optical flow is applied over con- secutive frames with the constant brightness assumption. In geometric matching [9, 48], apart from the encoder encodes source and support frames into feature maps, there exist transformer blocks which correlate two-frame fea- tures, e.g., the LoFTR module [48]. Since these network components consume two-frame inputs, the monocular pre- training task, e.g., the image classification and masked im- age modeling (MIM) defined on ImageNet dataset, is un- able to benefit the network. This limits both the geometric matching performance and its generalization capability. To address this, we reformulate the MIM from single masked image reconstruction to paired masked images re- construction, i.e., pMIM. Paired MIM benefits the geomet- ric matching as both tasks rely on the cross-frame module to correlate two frames inputs for prediction. This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 21909 With a pretrained encoder, the decoder in dense geomet- ric matching is still randomly initialized. Following the idea of pretraining encoder, we extend pMIM pretraining to the decoder. As part functionality of decoder is to upsample the coarse-scale initial prediction to the same resolution as input, we also task the decoder in pMIM to upsample the coarse-scale reconstruction to its original resolution. Corre- spondingly, we consist the decoder as stacks of the depth- wise convolution except for the last prediction head. With the depth-wise decoder, when transferring from pMIM to geometric matching, we duplicate the decoder along the channel dimension to finish the initialization. To this end, there exists only a small number of components in the de- coder randomly initialized, we pretrain the rest network components using synthetic image pair augmentation [54]. To further improve the dense geometric matching perfor- mance, we propose a cross-frame global matching module (CFGM). In CFGM, we first compute the correlation vol- ume. We model the correspondences of coarse scale pixels as a summation over the discrete coordinates in the support frame, weighted by the softmaxed correlation vector. How- ever, this modeling fails when multiple similar local patches exit. As a solution, we impose positional embeddings to the discrete coordinates and decode with a deep architec- ture to avoid ambiguity. Meanwhile, we notice that the tex- tureless surfaces are mostly planar structures described by a low-dimensional 8degree-of-freedom (DoF) homography matrix. We thus design a homography loss to augment the learning of the low DoF planar prior. We summarize our contributions as follows: •We introduce the paired masked image modeling pretext task, pretraining both the encoder and decoder of a dense geometric matching network. •We propose a novel cross-frame global matching module that is robust to textureless local patches. Since the most textureless patches are planar structures, we augment their learning with a homography loss. •We outperform dense and sparse geometric matching methods on diverse datasets.
Zhao_Quality-Aware_Pre-Trained_Models_for_Blind_Image_Quality_Assessment_CVPR_2023	Abstract Blind image quality assessment (BIQA) aims to auto- matically evaluate the perceived quality of a single image, whose performance has been improved by deep learning- based methods in recent years. However, the paucity of la- beled data somewhat restrains deep learning-based BIQA methods from unleashing their full potential. In this pa- per, we propose to solve the problem by a pretext task customized for BIQA in a self-supervised learning manner, which enables learning representations from orders of mag- nitude more data. To constrain the learning process, we propose a quality-aware contrastive loss based on a simple assumption: the quality of patches from a distorted image should be similar, but vary from patches from the same im- age with different degradations and patches from different images. Further, we improve the existing degradation pro- cess and form a degradation space with the size of roughly 2×107. After pre-trained on ImageNet using our method, models are more sensitive to image quality and perform sig- nificantly better on downstream BIQA tasks. Experimental results show that our method obtains remarkable improve- ments on popular BIQA datasets.	1. Introduction With the arrival of the mobile internet era, billions of im- ages are generated, uploaded and shared on various social media platforms, including Twitter, TikTok, etc[30]. As an essential indicator, image quality can help these service providers filter and deliver high-quality images to users, thereby improving Quality of Experience. Therefore, huge efforts [12,20,22,24,52,73] have been devoted to establish- ing an image quality assessment (IQA) method consistent with human viewers. In real-world scenarios, there usually exists no access to the reference images and the quality of reference images is suspicious. Thus, blind IQA (BIQA) methods are more attractive and applicable, despite full- reference IQA has achieved prospective results [33]. †Equal contribution. Figure 1. The two images in the first row are sampled from BIQA dataset CLIVE [20]. Although they have the same semantic mean- ing, their perceptual qualities are quite different: their mean opin- ion scores (MOS) are 31.83 and 86.35. The second row shows modified versions of the first two images and their MOSs (rated by 7 people). After different operations, the quality dramatically changed, while semantic meaning remains unchanged. Recently, deep learning-based BIQA methods have made tremendous improvements on in-the-wild IQA bench- marks [17, 29, 79]. However, this problem is far from re- solved and is hindered by the paucity of labeled data [34]. The largest (by far) available BIQA dataset, FLIVE [79], contains nearly 40,000 real-world distorted images. By comparison, the very popular entry-level image recognition dataset, CIFAR-100 [35], contains 60,000 labeled images. Accordingly, existing BIQA datasets are too small to train deep learning-based models effectively. Researchers present several methods to tackle this chal- lenge. A straight-forward way is to sample local patches and assign the label of the whole image ( i.e., mean opinion score, MOS) to the patches [4, 31, 32, 38, 60, 89, 90]. How- ever, the perceived scores of local image patches tend to differ from the score of the entire image [79, 89]. Another common strategy is to leverage domain knowledge from large-scale datasets ( e.g., ImageNet [14]) for other com- puter vision tasks [6, 32]. Nevertheless, these pre-trained models can be sub-optimal for BIQA tasks: images with the same content share the same semantic label, whereas, their quality may be different (Fig. 1). Some researchers This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 22302 propose to train a model on synthetic images with artificial degradation and then regress the model onto small-scale tar- get BIQA datasets [43, 73, 89]. However, images generated by rather simple degradation process with limited distortion types/levels are far from authentic. Further, synthetic im- ages are regularly distorted from limited pristine images of high quality, so the image content itself only has a marginal effect on the image quality. Yet in real-world scenarios, the image quality is closely related to its content, due to view- ers’ preferences for divergent contents [38, 62]. Self-supervised learning (SSL) or unsupervised learning is another potential choice to overcome the problem of lack- ing adequate training data, for its ability to utilize an amount of unlabeled data. Such technique is proven to be effective in many common computer vision tasks [2]. However, dif- ferent from models for these tasks which mainly focus on high-level information, representations learned for BIQA should be sensitive to all kinds of low-level distortions and high-level contents, as well as interactions between them. There are few research focused on such area in the litera- ture, let alone a deep SSL designed for BIQA with state-of- the-art performance [44]. In this work, we propose a novel SSL mechanism that distinguishes between samples with different percep- tual qualities, generating Quality-aware Pre-Trained (QPT) models for downstream BIQA tasks. Specifically, we sup- pose the quality of patches from a distorted image should be similar, but vary from patches from different images ( i.e., content-based negative ) and the same image with different degradations ( i.e.,degradation-based negative ). Moreover, inspired by recent progress in image restoration, we intro- duce shuffle order [84], high-order [70] and a skip oper- ation to the image degradation process, to simulate real- world distortions. In this way, models pre-trained on Im- ageNet are expected to extract quality-aware features, and boost downstream BIQA performances. We summarize the contributions of this work as follows: • We design a more complex degradation process suit- able for BIQA. It not only considers a mixture of multiple degradation types, but also incorporates with shuffle order, high-order and a skip operation, result- ing in a much larger degradation space. For a dataset with a size of 107, our method can generate more than 2×1014possible pairs for contrastive learning. • To fully exploit the abundant information hidden be- neath such an amount of data, we propose a novel SSL framework to generate QPT models for BIQA based on MoCoV2 [8]. By carefully designing positive/negative samples and customizing quality-aware contrastive loss, our approach enables models to learn quality- aware information rather than regular semantic-aware representation from massive unlabeled images.• Extensive experiments on five BIQA benchmark datasets (sharing the same pre-trained weight of QPT) demonstrate that our proposed method significantly outperforms other counterparts, which indicates the ef- fectiveness and generalization ability of QPT. It is also worth noting that the proposed method can be easily integrated with current SOTA methods by replacing their pre-trained weights.
Zhu_Visual_Prompt_Multi-Modal_Tracking_CVPR_2023	Abstract Visible-modal object tracking gives rise to a series of downstream multi-modal tracking tributaries. To inherit the powerful representations of the foundation model, a natural modus operandi for multi-modal tracking is full ﬁne-tuning on the RGB-based parameters. Albeit effective, this man- ner is not optimal due to the scarcity of downstream data and poor transferability, etc. In this paper, inspired by the recent success of the prompt learning in language models, we develop Visual Prompt multi-modal Tracking (ViPT), which learns the modal-relevant prompts to adapt the frozen pre-trained foundation model to various downstream multi- modal tracking tasks. ViPT ﬁnds a better way to stimulate the knowledge of the RGB-based model that is pre-trained at scale, meanwhile only introducing a few trainable pa- rameters (less than 1% of model parameters). ViPT outper- forms the full ﬁne-tuning paradigm on multiple downstream tracking tasks including RGB+Depth, RGB+Thermal, and RGB+Event tracking. Extensive experiments show the po- tential of visual prompt learning for multi-modal tracking, and ViPT can achieve state-of-the-art performance while satisfying parameter efﬁciency. Code and models are avail- able at https://github.com/jiawen-zhu/ViPT.	1. Introduction RGB-based tracking, a foundation task of visual object tracking, gains from large-scale benchmarks [9, 13, 27, 36, 39, 45] provided by the community, and many excellent works [2, 3, 5, 7, 22] have spurted out over the past decades. Despite the promising results, object tracking based on pure RGB sequences is still prone to failure in some com- plex and corner scenarios, e.g., extreme illumination, back- ground clutter, and motion blur. Therefore, multi-modal tracking is drawing increasing attention due to the ability to achieve more robust tracking by utilizing inter-modal com- plementarity, among which RGB+Depth (RGB-D) [47,60], yEqual contribution. BCorresponding author: Dr. Dong Wang. (b) Multi -modal Tracking (existing)(a) RGB -based Tracking (c) ViPT (ours)Foundation Task Pre -training Full Fine -tuning Prompt- tuning FusionFeature ExtractingFoundation Model Prompts… …RGBRGB RGB Feature ExtractingFoundation ModelDownstream dataLarge -scale data Foundation modality Auxiliary modality EventThermalDepth EventThermalDepthFigure 1. Existing multi-modal tracking paradigm vs.ViPT. (a) Foundation model training on large-scale RGB sequences. (a)!(b) Existing multi-modal methods extend the off-the-shelf foundation model and conduct full ﬁne-tuning on downstream tracking tasks. (a)!(c) We investigate the design of a multi-modal tracking method in prompt-learning paradigm. Compared to (b), the proposed ViPT has a more concise network structure, bene- ﬁts from parameter-friendly prompt-tuning, and narrows the gap between the foundation and downstream models. RGB+Thermal (RGB-T) [44, 56], and RGB+Event (RGB- E) [51, 52] are represented. However, as the downstream task of RGB-based track- ing, the main issue encountered by multi-modal tracking is the lack of large-scale datasets. For example, the widely used RGB-based tracking datasets, GOT-10k [13], Track- ingNet [36], and LaSOT [9], contain 9.3K, 30.1K, and 1.1K sequences, corresponding to 1.4M, 14M, and 2.8M frames for training. Whereas the largest training datasets in multi-modal tracking, DepthTrack [47], LasHeR [25], VisEvent [43], contain 150, 979, 500 training sequences, corresponding to 0.22M, 0.51M, 0.21M annotated frame pairs, which is at least an order of magnitude less than the former. Accounting for the above limitation, multi-modal tracking methods [43, 47, 61] usually utilize pre-trained RGB-based trackers and perform ﬁne-tuning on their task- oriented training sets (as shown in Figure 1 (a) !(b)). DeT [47] adds a depth feature extraction branch to the orig- inal ATOM [7] or DiMP [3] tracker and ﬁne-tunes on RGB- D training data. Zhang et al. [57] extend SiamRPN++ [21] with dual-modal inputs for RGB-T tracking. They ﬁrst con- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 9516 struct a unimodal tracking network trained on RGB data, then tune the whole extended multi-modal network with RGB-T image pairs. Similarly, Wang et al. [43] develop dual-modal trackers by extending single-modal ones with various fusion strategies for visible and event ﬂows and per- form extra model training on the RGB-E sequences. Al- though effective, the task-oriented full-tuning approach has some drawbacks. (i) Full ﬁne-tuning the model is time ex- pensive and inefﬁcient, and the burden of parameters stor- ing is large, which is unfriendly to numerous applications and cumbersome to transfer deploying. (ii) Full ﬁne-tuning is unable to obtain generalized representation due to the limited annotated samples, the inability to utilize the pre- trained knowledge of the foundation model trained on large- scale datasets. Thus, a natural question is thrown up: Is there a more effective manner to adapt the RGB-based foun- dation model to downstream multi-modal tracking? More recently, in Natural Language Processing (NLP) ﬁeld, researchers have injected textual prompts into down- stream language models to effectively exploit the repre- sentational potential of foundation models, this method, is known as prompt-tuning. After that, a few researchers have tried to freeze the entire upstream model and add only some learnable parameters to the input side to learn valid visual prompts. Existing researches [1, 14, 38, 59] show its great potential and visual prompt learning is ex- pected to be an alternative to full ﬁne-tuning. Intuitively, there is a large inheritance between multi-modal and sin- gle RGB-modal tracking, which should share most of prior knowledge on feature extraction or attention patterns. In this spirit, we present ViPT , a uniﬁed visual prompt-tuning paradigm for downstream multi-modal tracking. Instead of fully ﬁne-tuning an RGB-based tracker combined with an auxiliary-modal branch, ViPT freezes the whole foun- dation model and only learns a few modal-speciﬁc visual prompts, which inherits the RGB-based model parame- ters trained at scale to the maximum extent (see Figure 1 (c)). Different from the prompt learning of other single- modal vision tasks, ViPT introduces additional auxiliary- modal inputs into the prompt-tuning process, adapting the foundation model to downstream tasks while simultane- ously learning the association between different modalities. Speciﬁcally, ViPT inserts several simple and lightweight modality-complementary prompter (MCP) blocks into the frozen foundation model to effectively learn the inter-modal complementarities. Notably, ViPT is a general framework for various downstream multi-modal tracking tasks, includ- ing RGB-D, RGB-T, and RGB-E tracking. We summarize the contribution of our work as follows: A visual prompt tracking framework is proposed to achieve task-oriented multi-modal tracking. Facili- tated by learned prompts, the off-the-shelf foundation model can be effectively adapted from RGB domainto downstream multi-modal tracking tasks. Besides, ViPT is a general method that can be applied to vari- ous tasks, i.e., RGB-D, RGB-T, and RGB-E tracking. A modality-complementary prompter is designed to generate valid visual prompts for the task-oriented multi-modal tracking. The auxiliary-modal inputs are streamlined to a small number of prompts instead of designing an extra network branch. Extensive experiments show that our method achieves SOTA performance on multiple downstream multi- modal tracking tasks while maintaining parameter- efﬁcient (<1% trainable parameters).
Zhu_ConQueR_Query_Contrast_Voxel-DETR_for_3D_Object_Detection_CVPR_2023	Abstract Although DETR-based 3D detectors simplify the detec- tion pipeline and achieve direct sparse predictions, their performance still lags behind dense detectors with post- processing for 3D object detection from point clouds. DE- TRs usually adopt a larger number of queries than GTs (e.g., 300 queries v.s. ∼40 objects in Waymo) in a scene, which inevitably incur many false positives during infer- ence. In this paper, we propose a simple yet effective sparse 3D detector, named QueryContrast Voxel-DET R(Con- QueR ), to eliminate the challenging false positives, and achieve more accurate and sparser predictions. We observe that most false positives are highly overlapping in local re- gions, caused by the lack of explicit supervision to discrimi- nate locally similar queries. We thus propose a Query Con- trast mechanism to explicitly enhance queries towards their best-matched GTs over all unmatched query predictions. This is achieved by the construction of positive and negative GT-query pairs for each GT, and a contrastive loss to en- hance positive GT-query pairs against negative ones based on feature similarities. ConQueR closes the gap of sparse and dense 3D detectors, and reduces ∼60% false positives. Our single-frame ConQueR achieves 71.6 mAPH/L2 on the challenging Waymo Open Dataset validation set, outper- forming previous sota methods by over 2.0mAPH/L2. Code	1. Introduction 3D object detection from point clouds has received much attention in recent years [7, 32, 34, 47, 52] as its wide ap- plications in autonomous driving, robots navigation, etc. State-of-the-art 3D detectors [7, 31, 33, 53] still adopt dense predictions with post-processing ( e.g., NMS [2]) to obtain final sparse detections. This indirect pipeline usually in- volves many hand-crafted components (e.g., anchors, center masks) based on human experience, which involves much effort for tuning, and prevents dense detectors from being (a) Voxel-DETR(b) ConQueRFigure 1. Comparison of our baseline V oxel-DETR and ConQueR. GTs (green) and predictions (blue) of an example scene in the WOD is visualized. Sparse predictions of V oxel-DETR still con- tain many highly overlapped false positives (in the red dashed cir- cle), while ConQueR can generate much sparser predictions. optimized end-to-end to achieve optimal performance. Re- cently, DETR-based 2D detectors [3, 39, 49, 57] show that transformers with direct sparse predictions can greatly sim- plify the detection pipeline, and lead to better performance. However, although many efforts [1, 26, 27] have been made towards direct sparse predictions for 3D object detection, because of the different characteristics of images and point clouds ( i.e., dense and ordered images v.s.sparse and irreg- ular points clouds), performance of sparse 3D object detec- tors still largely lags behind state-of-the-art dense detectors. To achieve direct sparse predictions, DETRs usually adopt a set of object queries [1, 3, 27, 39, 49, 57], and re- sort to the one-to-one Hungarian Matching [17] to assign ground-truths (GTs) to object queries. However, to guaran- tee a high recall rate, those detectors need to impose much more queries than the actual number of objects in a scene. For example, recent works [1, 27] select top- 300 scored query predictions to cover only ∼40 objects in each scene of Waymo Open Dataset (WOD) [36], while 2D DETR de- tectors [3,39,49,57] use 10 ×more predictions than the av- erage GT number of MS COCO [22]. As shown in Fig. 1(a), we visualize an example scene by a baseline DETR-based This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 9296 3D detector, named V oxel-DETR, which shows its top- 300 scored predictions. Objects are generally small and densely populated in autonomous driving scenes, while 3D DETRs adopt the same fixed top- Nscored predictions as 2D DE- TRs, and lack a mechanism to handle such small and dense objects. Consequently, they tend to generate densely over- lapped false positives (in the red-dashed circle), harming both the accuracy and sparsity [29, 39] of final predictions. We argue the key reason is that the Hungarian Match- ing in existing 3D DETRs only assigns each GT to its best matched query, while all other unmatched queries near this GT are not effectively suppressed. For each GT, the one- to-one matching loss solely forces all unmatched queries to predict the same “no-object” label, and the best matched query are supervised without considering its relative rank- ing to its surrounding unmatched queries. This design causes the detectors to be insufficiently supervised in dis- criminating similar query predictions for each GT, leading to duplicated false positives for scenes with densely popu- lated objects. To overcome the limitations of current supervision, we introduce a simple yet novel Query Contrast strategy to ex- plicitly suppress predictions of all unmatched queries for each GT, and simultaneously enhance the best matched query to generate more accurate predictions in a contrastive manner. The Query Contrast strategy is integrated into our baseline V oxel-DETR, which consists of a sparse 3D con- volution backbone to extract features from voxel grids, and a transformer encoder-decoder architecture with a bipar- tite matching loss to directly generate sparse predictions. Our Query Contrast mechanism involves the construction of positive and negative GT-query pairs, and the contrastive learning on all GT-query pairs to supervise both matched and unmatched queries with knowledge of the states of their surrounding queries. Such GT-query pairs are directly cre- ated by reusing the Hungarian Matching results: each GT and its best matched query form the positive pair, and all other unmatched queries of the same GT then form nega- tive pairs. To quantitively measure the similarities of the GT-query pairs, we formulate the object queries to be the same as GT boxes ( i.e., using only box categories, loca- tions, sizes and orientations), such that GTs and object queries can be processed by the same transformer decoder, and embedded into a unified feature space to properly cal- culate their similarities. Given the GT-query similarities, we adopt the contrastive learning loss [5, 12, 54] to effec- tively enhance the positive (matched) query’s prediction for each GT, and suppress those of all its negative queries at the same time. Moreover, to further improve the contrastive supervision, we construct multiple positive GT-query pairs for each GT by adding small random noises to the original GTs, which greatly boost the training efficiency and effec- tiveness. The resulting sparse 3D detector QueryContrastV oxel-DET R(ConQueR ) significantly improves the detec- tion performance and sparsity of predictions, as shown in Fig. 1(b). Besides, ConQueR abandons the fixed top- Npre- diction scheme and achieves dynamic prediction numbers across scenes. ConQueR reduces ∼60% false positives and sets new records on the challenging Waymo Open Dataset (WOD) [36]. Contributions are summarized as bellow: 1. We introduce a novel Query Contrast strategy into DETR-based 3D detectors to effectively eliminate densely overlapped false positives and achieve more accurate predictions. 2. We propose to construct multi-positive contrastive training, which greatly improve the effectiveness and efficiency of our Query Contrast mechanism. 3. Our proposed sparse 3D detector ConQueR closes the gap between sparse and dense 3D detectors, and sets new records on the challenging WOD benchmark.
Zheng_EXIF_As_Language_Learning_Cross-Modal_Associations_Between_Images_and_Camera_CVPR_2023	Abstract We learn a visual representation that captures informa- tion about the camera that recorded a given photo. To do this, we train a multimodal embedding between image patches and the EXIF metadata that cameras automatically insert into image files. Our model represents this meta- data by simply converting it to text and then processing it with a transformer. The features that we learn significantly outperform other self-supervised and supervised features on downstream image forensics and calibration tasks. In particular, we successfully localize spliced image regions “zero shot” by clustering the visual embeddings for all of the patches within an image.	1. Introduction A major goal of the computer vision community has been to use cross-modal associations to learn concepts that would be hard to glean from images alone [2]. A particular focus has been on learning high level semantics, such asobjects, from other rich sensory signals, like language and sound [58, 62]. By design, the representations learned by these approaches typically discard imaging properties, such as the type of camera that shot the photo, its lens, and the exposure settings, which are not useful for their cross-modal prediction tasks [17]. We argue that obtaining a complete understanding of an image requires both capabilities — for our models to per- ceive not only the semantic content of a scene, but also the properties of the camera that captured it. This type of low level understanding has proven crucial for a variety of tasks, from image forensics [33, 52, 80] to 3D reconstruc- tion [34, 35], yet it has not typically been a focus of rep- resentation learning. It is also widely used in image gen- eration, such as when users of text-to-image tools specify camera properties with phrases like “DSLR photo” [59,63]. We propose to learn low level imaging properties from the abundantly available (but often neglected) camera meta- data that is added to the image file at the moment of cap- ture. This metadata is typically represented as dozens of Exchangeable Image File Format (EXIF) tags that describe the camera, its settings, and postprocessing operations that This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 6945 were applied to the image: e.g., Model : “iPhone 4s ” orFocal Length : “35.0 mm ”. We train a joint embed- ding through contrastive learning that puts image patches into correspondence with camera metadata (Fig. 1a). Our model processes the metadata with a transformer [75] after converting it to a language-like representation. To do this conversion, we take advantage of the fact that EXIF tags are typically stored in a human-readable (and text-based) format. We convert each tag to text, and then concate- nate them together. Our model thus closely resembles con- trastive vision-and-language models, such as CLIP [62], but with EXIF-derived text in place of natural language. We show that our model can successfully estimate cam- era properties solely from images, and that it provides a useful representation for a variety of image forensics and camera calibration tasks. Our approaches to these tasks do not require camera metadata at test time. Instead, camera properties are estimated implicitly from image content via multimodal embeddings. We evaluate the learned feature of our model on two clas- sification tasks that benefit from a low-level understanding of images: estimating an image’s radial distortion param- eter, and distinguishing real and manipulated images. We find that our features significantly outperform alternative supervised and self-supervised feature sets. We also show that our embeddings can be used to de- tect image splicing “zero shot” (i.e., without labeled data), drawing on recent work [8, 33, 54] that detects inconsisten- cies in camera fingerprints hidden within image patches. Spliced images contain content from multiple real images, each potentially captured with a different camera and imag- ing pipeline. Thus, the embeddings that our model assigns to their patches, which convey camera properties, will have less consistency than those of real images. We detect ma- nipulations by flagging images whose patch embeddings do not fit into a single, compact cluster. We also localize spliced regions by clustering the embeddings within an im- age (Fig. 1b). We show through our experiments that: • Camera metadata provides supervision for self- supervised representation learning. • Image patches can be successfully associated with camera metadata via joint embeddings. • Image-metadata embeddings are a useful representation for forensics and camera understanding tasks. • Image manipulations can be identified “zero shot” by identifying inconsistencies in patch embeddings.
Zhou_Query-Centric_Trajectory_Prediction_CVPR_2023	Abstract Predicting the future trajectories of surrounding agents is essential for autonomous vehicles to operate safely. This paper presents QCNet, a modeling framework toward push- ing the boundaries of trajectory prediction. First, we iden- tify that the agent-centric modeling scheme used by existing approaches requires re-normalizing and re-encoding the in- put whenever the observation window slides forward, lead- ing to redundant computations during online prediction. To overcome this limitation and achieve faster inference, we introduce a query-centric paradigm for scene encoding, which enables the reuse of past computations by learning representations independent of the global spacetime coordi- nate system. Sharing the invariant scene features among all target agents further allows the parallelism of multi-agent trajectory decoding. Second, even given rich encodings of the scene, existing decoding strategies struggle to capture the multimodality inherent in agents’ future behavior, espe- cially when the prediction horizon is long. To tackle this challenge, we first employ anchor-free queries to generate trajectory proposals in a recurrent fashion, which allows the model to utilize different scene contexts when decod- ing waypoints at different horizons. A refinement module then takes the trajectory proposals as anchors and leverages anchor-based queries to refine the trajectories further. By supplying adaptive and high-quality anchors to the refine- ment module, our query-based decoder can better deal with the multimodality in the output of trajectory prediction. Our approach ranks 1ston Argoverse 1 and Argoverse 2 motion forecasting benchmarks, outperforming all methods on all main metrics by a large margin. Meanwhile, our model can achieve streaming scene encoding and parallel multi-agent decoding thanks to the query-centric design ethos.	1. Introduction Making safe decisions for autonomous vehicles requires accurate predictions of surrounding agents’ future trajec- tories. In recent years, learning-based methods have been observation @ 𝑡observation @ 𝑡+1 agent history reference frame of agent statereference frame of map polygonFigure 1. Illustration of our query-centric reference frame , where we build a local coordinate system for each spatial-temporal element, including map polygons and agent states at all time steps. In the attention-based encoder, all scene elements’ queries are de- rived and updated in their local reference frames. widely used for trajectory prediction [14, 31, 37, 38, 46, 56]. Despite the considerable efforts made to enhance models’ forecasting ability, there is still a long way to go before fully addressing the problem of trajectory prediction. Why is this task so challenging, and what inability lies in existing ap- proaches? We attempt to answer these questions from the following two perspectives: (i)While the flourishing forecasting models have achieved impressive performance on trajectory prediction benchmarks [7,13,49], today’s most advanced architectures specialized for this task [37, 38, 46, 56] fail to process the heterogeneous traffic scenes efficiently . In an autonomous driving system, data frames arrive at the prediction module sequentially as a stream of sparse scene context, including the high-definition vector map and the surrounding agents’ kinematic states. A model must learn expressive representa- tions of these scene elements to achieve accurate forecasts. With the continuing development of modeling techniques for sparse context encoding [14, 31, 50], the research com- munity has witnessed rapid progress toward more powerful trajectory predictors. Notably, factorized attention-based Transformers [37,38,56] have recently raised prediction ac- curacy to an unprecedented level. However, they require learning attention-based representations for each spatial- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 17863 temporal scene element and suffer from prohibitively high costs when processing dense traffic scenes. As every mini- mal delay may lead to catastrophic accidents in autonomous driving, the unmet need for real-time predictions has limited the applicability of state-of-the-art approaches. (ii)The immense uncertainty in the output of trajectory prediction, which grows explosively as the prediction hori- zon lengthens, has troubled the research community con- stantly. For example, a vehicle at an intersection may turn or go straight depending on the driver’s long-term goal. To avoid missing any potential behavior, a model must learn to capture the underlying multimodal distribution rather than simply predicting the most frequent mode. This learning task is challenging since only one possibility is logged in each training sample. To ease the learning difficulty, a body of works utilizes handcrafted anchors as guidance for mul- timodal prediction [6, 12, 39, 53, 55]. Their effectiveness, however, is subject to the quality of the anchors. Typically, these methods fail to work well when few anchors can pre- cisely cover the ground truth. This problem is exacerbated in long-term prediction, where the search space for anchors is much larger. Some other works [10, 31, 38, 46, 56] cir- cumvent this issue by directly predicting multiple trajecto- ries, albeit at the risk of mode collapse and training instabil- ity [33, 41]. Due to the lack of spatial priors, these methods also fail to produce accurate long-term forecasts. The analysis above drives us to propose a trajectory pre- diction framework, termed as QCNet, to overcome the lim- itations of previous solutions. First , we note that it is possi- ble to achieve faster online inference while also benefiting from the power of factorized attention, but the agent-centric encoding scheme [25, 27, 46, 56] used by existing methods serves as an impediment. Each time a new data frame ar- rives, the observation window slides one step forward and overlaps with its predecessor substantially, which provides opportunities for models to reuse the previously computed encodings. However, agent-centric approaches require nor- malizing the input based on the latest agent states’ positions, necessitating the re-encoding of scene elements whenever the observation window slides forward. To address this is- sue, we introduce a query-centric paradigm for scene en- coding (see Fig. 1). The crux of our design ethos lies in processing all scene elements in their local spacetime ref- erence frames and learning representations independent of the global coordinates. This strategy enables us to cache and reuse the previously computed encodings, spreading the computation across all observation windows and thereby re- ducing inference latency. The invariant scene features can also be shared among all target agents in the scene to en- able the parallelism of multi-agent decoding. Second , to better utilize the scene encodings for multimodal and long- term prediction, we use anchor-free queries to retrieve the scene context recurrently and let them decode a short seg-ment of future waypoints at each recurrence. This recurrent mechanism eases the modeling burden on the queries by al- lowing them to focus on different scene contexts when pre- dicting waypoints at different horizons. The high-quality trajectories predicted by the recurrent decoder serve as dy- namic anchors in the subsequent refinement module, where we use anchor-based queries to refine the trajectory propos- als based on the scene context. As a result, our query-based decoding pipeline incorporates the flexibility of anchor-free methods into anchor-based solutions, taking the best of both worlds to facilitate multimodal and long-term prediction. Our proposed query-centric encoding paradigm is the first that can exploit the sequential nature of trajectory pre- diction to achieve fast online inference. Besides, our query- based decoder exhibits superior performance for multi- modal and long-term prediction. Experiments show that our approach achieves state-of-the-art results, ranking 1ston two large-scale motion forecasting benchmarks [7, 49].
Zhu_BiFormer_Vision_Transformer_With_Bi-Level_Routing_Attention_CVPR_2023	Abstract As the core building block of vision transformers, atten- tion is a powerful tool to capture long-range dependency. However, such power comes at a cost: it incurs a huge computation burden and heavy memory footprint as pair- wise token interaction across all spatial locations is com- puted. A series of works attempt to alleviate this problem by introducing handcrafted and content-agnostic sparsity into attention, such as restricting the attention operation to be inside local windows, axial stripes, or dilated windows. In contrast to these approaches, we propose a novel dy- namic sparse attention via bi-level routing to enable a more ﬂexible allocation of computations with content awareness. Speciﬁcally, for a query, irrelevant key-value pairs are ﬁrst ﬁltered out at a coarse region level, and then ﬁne-grained token-to-token attention is applied in the union of remain- ing candidate regions ( i.e., routed regions). We provide a simple yet effective implementation of the proposed bi- level routing attention, which utilizes the sparsity to save both computation and memory while involving only GPU- friendly dense matrix multiplications. Built with the pro- posed bi-level routing attention, a new general vision trans- former, named BiFormer, is then presented. As BiFormer attends to a small subset of relevant tokens in a query adap- tivemanner without distraction from other irrelevant ones, it enjoys both good performance and high computational efﬁciency, especially in dense prediction tasks. Empirical results across several computer vision tasks such as image classiﬁcation, object detection, and semantic segmentation verify the effectiveness of our design. Code is available at https://github.com/rayleizhu/BiFormer .	1. Introduction Transformer has many properties that are suitable for building powerful data-driven models. First, it is able to capture long-range dependency in the data [27,40]. Second, yCorresponding author.it is almost inductive-bias-free and thus makes the model more ﬂexible to ﬁt tons of data [14]. Last but not least, it enjoys high parallelism, which beneﬁts training and infer- ence of large models [12, 31, 34, 40]. Hence, transformer has not only revolutionized natural language processing but also shown very promising progress in computer vision. The computer vision community has witnessed an explo- sion of vision transformers in the past two years [1, 13, 14, 27, 42, 44]. Among these works, a popular topic is to im- prove the core building block, i.e., attention. In contrast to convolution, which is intrinsically a local operator, a cru- cial property of attention is the global receptive ﬁeld, which empowers vision transformers to capture long-range depen- dency [40]. However, such a property comes at a cost: as attention computes pairwise token afﬁnity across all spatial locations, it has a high computational complexity and incurs heavy memory footprints. To alleviate the problem, a promising direction is to in- troduce sparse attention [5] to vision transformers, so that each query attends to a small portion of key-value pairs instead of all. In this fashion, several handcrafted sparse patterns have been explored, such as restricting attention in local windows [27], dilated windows [39, 44], or axial stripes [44]. On the other hand, there are also works try- ing to make the sparsity adaptive to data [4, 45]. However, while they use different strategies to merge or select key/- value tokens, these tokens are query-agnostic, i.e., they are shared by all queries. Nonetheless, according to the visual- ization of pretrained ViT1[14] and DETR2[1], queries in different semantic regions actually attend to quite different key-value pairs. Hence, forcing all queries to attend to the same set of tokens may be suboptimal. In this paper, we seek an attention mechanism with dy- namic, query-aware sparsity. Basically, we aim for each query to attend to a small portion of the most semantically relevant key-value pairs. The ﬁrst problem comes as how 1https://epfml.github.io/attention-cnn/ 2https : / / colab . research . google . com / github / facebookresearch/detr/blob/colab/notebooks/detr_ attention.ipynb This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 10323 (a) V anilla Attention (b) Local Attentionquery key/value local window (c) Axial Attention (d) Dilated Attention (e) Deformable Attention (f) Bi-level Routing AttentionFigure 1. Vanilla attention and its sparse variants. (a)Vanilla attention operates gloabally and incurs high computational complexity and heavy memory footprint. (b)-(d) Several works attempt to alleviate the complexity by introducing sparse attention with different handcrafted patterns, such as local window [27, 44], axial stripe [13], dilated window [39, 44]. (e)Deformable attention [45] enables image-adaptive sparsity via deforming a regular grid. (f)We achieve dynamic, query-aware sparsity with bi-level routing attention, which ﬁrst searches top- k(k= 3in this case) relevant regions, and then attends to the union of them. to locate these key-value pairs to attend. For example, if we select key-value pairs in a per-query manner as done in [16], it still requires evaluation of pairwise afﬁnity between all queries and keys, and hence has the same complexity of vanilla attention. Another possibility is to predict attention offsets based on local context for each query [9, 45], and hence pairwise afﬁnity computation is avoided. However, in this way, it is problematic to model long-range depen- dency [45]. To locate valuable key-value pairs to attend globally with high efﬁciency, we propose a region-to-region routing ap- proach. Our core idea is to ﬁlter out the most irrelevant key-value pairs at a coarse-grained region level, instead of directly at the ﬁne-grained token level. This is done by ﬁrst constructing a region-level afﬁnity graph and then pruning it to keep only top- kconnections for each node. Hence, each region only needs to attend to the top- krouted regions. With the attending regions determined, the next step is to apply token-to-token attention, which is non-trivial as key-value pairs are now assumed to be spatially scattered. For this case, while the sparse matrix multiplication is applicable, it is inefﬁcient in modern GPUs, which rely on coalesced memory operations, i.e., accessing blocks of dozens of con- tiguous bytes at once [29]. Instead, we propose a simple so-lution via gathering key/value tokens, where only hardware- friendly dense matrix multiplications are involved. We refer to this approach as Bi-level Routing Attention (BRA), as it contains a region-level routing step and a token-level atten- tion step. By using BRA as the core building block, we propose BiFormer, a general vision transformer backbone that can be used for many applications such as classiﬁcation, object detection, and semantic segmentation. As BRA enables Bi- Former to attend to a small subset of the most relevant key/- value tokens for each query in a content-aware manner, our model achieves a better computation-performance trade-off. For example, with 4.6G FLOPs computation, BiFormer-T achieves 83.8% top-1 accuracy on ImageNet-1K classiﬁ- cation, which is the best as far as we know under similar computation budgets without training with external data or distillation [22,38]. The improvements are also consistently shown in downstream tasks such as instance segmentation and semantic segmentation. To summarize, our contributions are as follows. We in- troduce a novel bi-level routing mechanism to vanilla at- tention, which enables content-aware sparse patterns in a query-adaptive manner. Using the bi-level routing atten- tion as the basic building block, we propose a general vi- 10324 sion transformer named BiFormer. Experimental results on various computer vision tasks including image classiﬁca- tion, object detection, and semantic segmentation show that the proposed BiFormer achieves signiﬁcantly better perfor- mances over the baselines under similar model sizes.
Zhao_Search-Map-Search_A_Frame_Selection_Paradigm_for_Action_Recognition_CVPR_2023	Abstract Despite the success of deep learning in video under- standing tasks, processing every frame in a video is com- putationally expensive and often unnecessary in real-time applications. Frame selection aims to extract the most in- formative and representative frames to help a model better understand video content. Existing frame selection methods either individually sample frames based on per-frame im- portance prediction, without considering interaction among frames, or adopt reinforcement learning agents to find rep- resentative frames in succession, which are costly to train and may lead to potential stability issues. To overcome the limitations of existing methods, we propose a Search-Map- Search learning paradigm which combines the advantages of heuristic search and supervised learning to select the best combination of frames from a video as one entity. By com- bining search with learning, the proposed method can bet- ter capture frame interactions while incurring a low infer- ence overhead. Specifically, we first propose a hierarchical search method conducted on each training video to search for the optimal combination of frames with the lowest er- ror on the downstream task. A feature mapping function is then learned to map the frames of a video to the representa- tion of its target optimal frame combination. During infer- ence, another search is performed on an unseen video to se- lect a combination of frames whose feature representation is close to the projected feature representation. Extensive ex- periments based on several action recognition benchmarks demonstrate that our frame selection method effectively im- proves performance of action recognition models, and sig- nificantly outperforms a number of competitive baselines.	1. Introduction Videos have proliferated online in recent years with the popularity of social media, and have become a major form of content consumption on the Internet. The abundant video data has greatly encouraged the development of deep learn- ing techniques for video content understanding. As one of the most important tasks, action recognition aims to iden- tify relevant actions described in videos, and plays a vital role to other downstream tasks like video retrieval and rec- ommendation. Due to the high computational cost of processing frames in a video, common practices of action recognition involve sampling a subset of frames or clips uniformly [31] or densely [26, 37] from a given video a serve as the input to a content understanding model. However, since frames in a video may contain redundant information and are not equally important, simple sampling methods are often inca- pable of capturing such knowledge and hence can lead to sub-optimal action recognition results. Prior studies attempt to actively select relevant video frames to overcome the limitation of straightforward sam- pling, achieving improvements to model performance. Heuristic methods are proposed to rank and select frames according to the importance score of each frame/clip cal- culated by per-frame prediction [15, 19]. Despite the effec- tiveness, these methods heavily rely on per-frame features, without considering the interaction or diversity among se- lected frames. Reinforcement learning (RL) has also been proposed to identify informative frames by formulating frame selection as a Markov decision process (MDP) [8, 9, 33,35,36]. However, existing RL-based methods may suffer from training stability issues and rely on a massive amount of training samples. Moreover, RL methods make an MDP assumption that frames are selected sequentially depending This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 10627 on observations of already selected frames, and thus cannot adjust prior selections based on new observations. In this work, we propose a new learning paradigm named Search-Map-Search (SMS), which directly searches for the best combination of frames from a video as one entity. SMS formulates the problem of frame selection from the perspec- tive of heuristic search in a large space of video frame com- binations, which is further coupled with a learnable map- ping function to generalize to new videos and achieve effi- cient inference. Specifically, we propose a hierarchical search algorithm to efficiently find the most favorable frame combinations on training videos, which are then used as explicit supervision information to train a feature mapping function that maps the feature vectors of an input video to the feature vector of the desirable optimal frame combination. During inference on an unseen query video, the learned mapping function projects the query video onto a target feature vector for the desired frame combination, where another search process retrieves the actual frame combination that approximates the target feature vector. By combining search with learn- ing, the proposed SMS method can better capture frame in- teractions while incurring a low inference cost. The effectiveness of SMS is extensively evaluated on both the long untrimmed action recognition benchmarks, i.e., ActivityNet [2] and FCVID [18], and the short trimmed UCF101 task [27]. Experimental results show that SMS can significantly improve action recognition models and precisely recognize and produce effective frame selections. Furthermore, SMS significantly outperforms a range of other existing frame selection methods for the same num- ber of frames selected, while can still generate performance higher than existing methods using only 10% of all labeled video samples for training.
Zhao_CDDFuse_Correlation-Driven_Dual-Branch_Feature_Decomposition_for_Multi-Modality_Image_Fusion_CVPR_2023	Abstract Multi-modality (MM) image fusion aims to render fused images that maintain the merits of different modalities, e.g., functional highlight and detailed textures. To tackle the chal- lenge in modeling cross-modality features and decomposing desirable modality-speciﬁc and modality-shared features, we propose a novel Correlation-Driven feature Decompo- sition Fusion ( CDDFuse ) network. Firstly, CDDFuse uses Restormer blocks to extract cross-modality shallow features. We then introduce a dual-branch Transformer-CNN feature extractor with Lite Transformer (LT) blocks leveraging long- range attention to handle low-frequency global features and Invertible Neural Networks (INN) blocks focusing on ex- tracting high-frequency local information. A correlation- driven loss is further proposed to make the low-frequency features correlated while the high-frequency features un- correlated based on the embedded information. Then, the LT-based global fusion and INN-based local fusion layers output the fused image. Extensive experiments demonstrate that our CDDFuse achieves promising results in multiple fusion tasks, including infrared-visible image fusion and medical image fusion. We also show that CDDFuse can boost the performance in downstream infrared-visible se- mantic segmentation and object detection in a uniﬁed bench- mark. The code is available at https://github.com/ Zhaozixiang1228/MMIF-CDDFuse .	1. Introduction Image fusion is a basic image processing topic that aims to generate informative fused images by combining the im- portant information from source ones [47,75,78,79,87]. The fusion targets include digital [45,84], multi-modal [36,70,88] *Corresponding author. (a) Existing MMIF methods vs. CDDFuse. The base anddetail encoders are respon- sible for extracting global and local features, respectively. (b) CDDFuse (highlighted in yellow) achieves the state-of-the-art performance on MSRS [57] and RoadScene [71] in eight metrics. Figure 1. Workﬂow and performance comparison of our proposed CDDFuse with existing MM image fusion approaches. and remote sensing [4, 76, 91] images, etc. The Infrared- Visible image Fusion (IVF) and Medical Image Fusion (MIF) are two challenging sub-categories of Multi-Modality Image Fusion (MMIF), focusing on modeling the cross-modality features from all the sensors and aggregating them into the output images. Speciﬁcally, IVF targets fused images that preserve thermal radiation information in the input infrared images and detailed texture information in the input visible images. The fused images can avoid the shortcomings of visible images being sensitive to illumination conditions as well as the infrared images being noisy and low-resolution. Downstream recognition tasks, e.g., multi-modal saliency detection [33, 52, 63], object detection [6, 34, 58] and se- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 5906 mantic segmentation [37, 49 –51] can then beneﬁt from the obtained clearer representations of scenes and objects in IVF images. Similarly, MIF aims to clearly exhibit the abnor- malities by fusing multiple medical imaging modalities to reveal comprehensive information to assist diagnosis and treatment [21]. Many methods have been developed to tackle the MMIF challenges in recent years [35, 41, 44, 55, 73, 74, 82]. A common pipeline that demonstrated promising results uti- lizes CNN-based feature extraction and reconstruction in an Auto-Encoder (AE) manner [28, 29, 32, 88]. The workﬂow is illustrated in Fig. 1a. However, existing methods have three main shortcomings. First, the internal working mecha- nism of CNNs is difﬁcult to control and interpret, causing insufﬁcient extraction of cross-modality features. For ex- ample, in Fig. 1a, shared encoders in (I) and (II) cannot distinguish modality-speciﬁc features, while the private en- coders in (III) ignore features shared by modalities. Second, the context-independent CNN only extracts local informa- tion in a relatively small receptive ﬁeld, which can hardly extract global information for generating high-quality fused images [31]. Thus, it is still unclear whether the inductive biases of CNN are capable enough to extract features for all modalities. Third, the forward propagation of fusion networks often causes the loss of high-frequency informa- tion [42,89]. Our work explores a more reasonable paradigm to tackle the challenges in feature extraction and fusion. First, we aim to add correlation restrictions to the ex- tracted features and limit the solution space, which improves the controllability and interpretability of feature extraction. Our assumption is that, in the MMIF task, the input fea- tures of the two modalities are correlated at low frequen- cies, representing the modality-shared information, while the high-frequency feature is irrelevant and represents the unique characteristics of the respective modalities. Taking IVF as an example, since infrared and visible images come from the same scene, the low-frequency information of the two modalities contains statistical co-occurrences, such as background and large-scale environmental features. On the contrary, the high-frequency information of the two modali- ties is independent, e.g., the texture and detail information in the visible image and the thermal radiation information in the infrared image. Therefore, we aim to facilitate the extraction of modality-speciﬁc and modality-shared features by increas- ing and decreasing the correlation between low-frequency and high-frequency features, respectively. Second, from the architectural perspective, Vision Trans- formers [14, 38, 80] recently shows impressive results in computer vision, with self-attention mechanism and global feature extraction. However, Transformer-based methods are computationally expensive, which leaves room for further improvement with considering the efﬁciency-performance tradeoff of image fusion architectures. Therefore, we pro-pose integrating the advantages of local context extraction and computational efﬁciency in CNN and the advantages of global attention and long-range dependency modeling in Transformer to complete the MMIF task. Third, to solve the challenge of losing wanted high- frequency input information, we adopt the building block of Invertible Neural networks (INN) [13]. INN was proposed with invertibility by design, which prevents information loss through the mutual generation of input and output features and aligns with our goal of preserving high-frequency fea- tures in the fused images. To this end, we proposed the Correlation-Driven feature Decomposition Fusion (CDDFuse ) model, where modality- speciﬁc and modality-shared feature extractions are realized by a dual-branch encoder, with the fused image reconstructed by the decoder. The workﬂow is shown in Figs. 1a and 2. Our contributions can be summarized in four aspects: •We propose a dual-branch Transformer-CNN frame- work for extracting and fusing global and local features, which better reﬂects the distinct modality-speciﬁc and modality-shared features. •We reﬁne the CNN and Transformer blocks for a better adaptation to the MMIF task. Speciﬁcally, we are the ﬁrst to utilize the INN blocks for lossless information transmission and the LT blocks for trading-off fusion quality and computational cost. •We propose a correlation-driven decomposition loss function to enforce the modality shared/speciﬁc feature decomposition, which makes the cross-modality base features correlated while decorrelates the detailed high- frequency features in different modalities. •Our method achieves leading image fusion performance for both IVF and MIF. We also present a uniﬁed mea- surement benchmark to justify how the IVF fusion im- ages facilitate downstream MM object detection and semantic segmentation tasks.
Zhu_Patch-Mix_Transformer_for_Unsupervised_Domain_Adaptation_A_Game_Perspective_CVPR_2023	Abstract Endeavors have been recently made to leverage the vi- sion transformer (ViT) for the challenging unsupervised domain adaptation (UDA) task. They typically adopt the cross-attention in ViT for direct domain alignment. However , as the performance of cross-attention highly relies on thequality of pseudo labels for targeted samples, it becomes less effective when the domain gap becomes large. We solve this problem from a game theory’s perspective with the pro- posed model dubbed as PMTrans , which bridges source and target domains with an intermediate domain. Speciﬁcally, we propose a novel ViT-based module called PatchMix that effectively builds up the intermediate domain, i.e., proba- bility distribution, by learning to sample patches from both domains based on the game-theoretical models. This way, it learns to mix the patches from the source and target do- mains to maximize the cross entropy (CE), while exploiting two semi-supervised mixup losses in the feature and labelspaces to minimize it. As such, we interpret the process of UDA as a min-max CE game with three players, including the feature extractor , classiﬁer , and PatchMix, to ﬁnd the Nash Equilibria. Moreover , we leverage attention maps from ViT to re-weight the label of each patch by its importance, making it possible to obtain more domain-discriminative feature representations. We conduct extensive experiments on four benchmark datasets, and the results show that PMTrans signiﬁcantly surpasses the ViT-based and CNN- based SoTA methods by +3.6% on Ofﬁce-Home, +1.4% on Ofﬁce-31, and +17.7% on DomainNet, respectively. https: //vlis2022.github.io/cvpr23/PMTrans	1. Introduction Convolutional neural networks (CNNs) have achieved tremendous success on numerous vision tasks; however, theystill suffer from the limited generalization capability to a new domain due to the domain shift problem [ 50]. Unsupervised domain adaptation (UDA) tackles this issue by transferring *These authors contributed equally to this work. †Corresponding Author. 20253035 40455055Accuracy Target¬ tasksclp inf pnt qdr rel skt Figure 1. The classiﬁcation accuracy of our PMTrans surpasses the SoTA methods by +17.7% on the most challenging DomainNet dataset. Note that the target tasks treat one domain of DomainNet as the target domain and the others as the source domains. knowledge from a labeled source domain to an unlabeled target domain [ 30]. A signiﬁcant line of solutions reduces the domain gap based on the category-level alignment which produces pseudo labels for the target samples, such as metriclearning [ 14,53], adversarial training [ 12,17,34], and optimal transport [ 44]. Furthermore, several works [ 11,36] explore the potential of ViT for the non-trivial UDA task. Recently, CDTrans [ 45] exploits the cross-attention in ViT for direct domain alignment, buttressed by the crafted pseudo labels for target samples. However, CDTrans has a distinct limita- tion: as the performance of cross-attention highly depends on the quality of pseudo labels, it becomes less effective when the domain gap becomes large. As shown in Fig. 1, due to the signiﬁcant gap between the domain qdr and the other domains, aligning distributions directly between them performs poorly. In this paper, we probe a new problem for UDA: how to smoothly bridge the source and target domains by con- structing an intermediate domain with an effective ViT - based solution? The intuition behind this is that, compared to direct aligning domains, alleviating the domain gap be- tween the intermediate and source/target domain can facili- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 3561 Figure 2. PMTrans builds up the intermediate domain ( green patches ) via a novel PatchMix module by learning to sample patches from the source ( blue patches ) and target ( pink patches ) domains. PatchMix tries to maximize the CE ( ↑) between the intermediate domain and source/target domain, while the feature extractor and classiﬁer try to minimize it ( ↓) for aligning domains. tate domain alignment. Accordingly, we propose a novel and effective method, called PMTrans (PatchMix Transformer) to construct the intermediate representations. Overall, PM- Trans interprets the process of domain alignment as a min- max cross entropy (CE) game with three players, i.e., the feature extractor, a classiﬁer, and a PatchMix module, to ﬁnd the Nash Equilibria. Importantly, the PatchMix module is proposed to effectively build up the intermediate domain, i.e., probability distribution, by learning to sample patches from both domains with weights generated from a learnable Beta distribution based on the game-theoretical models [ 1,3,28], as shown in Fig. 2. That is, we aim to learn to mix patches from two domains to maximize the CE between the inter- mediate domain and source/target domain. Moreover, two semi-supervised mixup losses in the feature and label spacesare proposed to minimize the CE. Interestingly, we conclude that the source and target domains are aligned if mixing the patch representations from two domains is equivalent to mixing the corresponding labels . Therefore, the domain discrepancy can be measured based on the CE between the mixed patches and mixed labels. Eventually, the three play- ers have no incentive to change their parameters to disturb CE, meaning the source and target domains are well aligned. Unlike existing mixup methods [ 38,47,49], our proposed PatchMix subtly learns to combine the element-wise global and local mixture by mixing patches from the source and tar- get domains for ViT-based UDA. Moreover, we leverage the class activation mapping (CAM) from ViT to allocate the se- mantic information to re-weight the label of each patch, thus enabling us to obtain more domain-discriminative features. We conduct experiments on four benchmark datasets, in- cluding Ofﬁce-31 [ 33], Ofﬁce-Home [ 40], VisDA-2017 [ 32], and DomainNet [ 31]. The results show that the performance of PMTrans signiﬁcantly surpasses that of the ViT-based [36,45,46] and CNN-based SoTA methods [ 18,29,35]b y+3.6% on Ofﬁce-Home, +1.4% on Ofﬁce-31, and +17.7% on DomainNet (See Fig. 1), respectively. Our main contributions are four-fold: ( I) We propose a novel ViT-based UDA framework, PMTrans, to effectively bridge source and target domains by constructing the inter- mediate domain. ( II) We propose PatchMix, a novel module to build up the intermediate domain via the game-theoretical models. ( III) We propose two semi-supervised mixup losses in the feature and label spaces to reduce CE in the min-max CE game. ( IV) Our PMTrans surpasses the prior methods by a large margin on three benchmark datasets.
Zhou_UDE_A_Unified_Driving_Engine_for_Human_Motion_Generation_CVPR_2023	Abstract Generating controllable and editable human motion se- quences is a key challenge in 3D Avatar generation. It has been labor-intensive to generate and animate human mo- tion for a long time until learning-based approaches have been developed and applied recently. However, these ap- proaches are still task-specific or modality-specific [1] [6] [5] [18]. In this paper, we propose “UDE”, the first uni- fied driving engine that enables generating human motion sequences from natural language or audio sequences (see Fig. 1). Specifically, UDE consists of the following key components: 1) a motion quantization module based on VQVAE that represents continuous motion sequence as dis- crete latent code [33], 2) a modality-agnostic transformer encoder [34] that learns to map modality-aware driving signals to a joint space, and 3) a unified token transformer (GPT-like [24]) network to predict the quantized latent code index in an auto-regressive manner. 4) a diffusion motion decoder that takes as input the motion tokens and decodes them into motion sequences with high diversity. We evaluate our method on HumanML3D [8] and AIST+ [19] bench- marks, and the experiment results demonstrate our method achieves state-of-the-art performance. Project website: https://zixiangzhou916.github.io/UDE/	1. Introduction Synthesizing realistic human motion sequences has been a pilar component in many real-world applications. It is labor-intensive and tedious, and requires professional skills to achieve the creation of one single piece of motion se- quence synthesis, making it hard to be democratized for broad content generations. Recently, the emergence of mo- tion capture and pose estimation [15] [38] [27] [36] have made it possible to synthesize human motion sequences from VTubers or source videos thanks to the advances of deep learning. Although these approaches have simplified the creation of motion sequences, actors or highly corre- lated videos are still necessary, thus limiting the scalability as well as the controllability. The development of multi-modal machine learning paves a new way to human motion synthesis [1] [6] [8] [12] [17] [2]. For example, natural language descriptions could be used to drive human motion sequences directly [1] [6] [8]. The language description is a straightforward representa- tion for human users to control the synthesis. It provides a semantic clue of what the synthesized motion sequence should look like, and the editing could be conducted by sim- ply changing the language description. Language, however, does not cover the full domain of human motion sequences. In terms of dancing motion synthesis, for example, the nat- ural language is not sufficient to describe the dance rhythm. This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 5632 For such scenarios, audio sequences are used as guidance to help motion synthesis, so the synthesized motion could match the music beat rhythmically and choreography style. However, these approaches are studied separately in prior works. In many real-world applications, the characters are likely to perform a complex motion sequence composed of both rhythmic dances from music and certain actions de- scribed by language, and smooth transition between mixed modality inputs becomes vital important. As a result, multi- modal motion consistency would become an urgent issue to solve if employed siloed modality-specific models. To address above mentioned problems, in this work, we propose a Unified Driving Engine ( UDE ) which unifies the human motion generation driven by natural language and music clip in one shared model. Our model consists of four key components. First, we train a codebook using VQ-V AE. For the codebook, each code represents a certain pattern of the motion sequence. Second, we introduce a Modality-Agnostic Transformer Encoder ( MATE ). It takes the input of different modalities and transforms them into sequential embedding in one joint space. The third com- ponent is a Unified Token Transformer ( UTT ). We feed it with sequential embedding obtained by MATE and pre- dict the motion token sequences in an auto-regressive man- ner. The fourth component is a Diffusion Motion Decoder (DMD ). Unlike recent modality-specific works [30] [37], ourDMD is modality-agnostic. Given the motion token se- quences, DMD decodes them to motion sequences in con- tinuous space by the reversed diffusion process. We summarize our contributions in four folds: 1) We model the continuous human motion generation problem as a discrete token prediction problem. 2) We unify the text- driven and audio-driven motion generation into one sin- gle unified model. By learning MATE , we can map in- put sequences of different modalities into joint space. Then we can predict motion tokens with UTT regardless of the modality of input. 3) We propose DMD to decode the mo- tion tokens to motion sequence. Compared to the decoder in VQ-V AE, which generates deterministic samples, our DMD can generate samples with high diversity. 4) We eval- uate our method extensively and the results suggest that our method outperforms existing methods in both text-driven and audio-driven scenarios. More importantly, our experi- ment also suggests that our UDE enables smooth transition between mixed modality inputs.
Zhou_Non-Contrastive_Learning_Meets_Language-Image_Pre-Training_CVPR_2023	Abstract Contrastive language-image pre-training (CLIP) serves as a de-facto standard to align images and texts. Nonethe- less, the loose correlation between images and texts of web- crawled data renders the contrastive objective data ineffi- cient and craving for a large training batch size. In this work, we explore the validity of non-contrastive language- image pre-training (nCLIP), and study whether nice proper- ties exhibited in visual self-supervised models can emerge. We empirically observe that the non-contrastive objective benefits representation learning while sufficiently underper- forming under zero-shot recognition. Based on the above study, we further introduce xCLIP, a multi-tasking frame- work combining CLIP and nCLIP, and show that nCLIP aids CLIP in enhancing feature semantics. The synergy between two objectives lets xCLIP enjoy the best of both worlds: superior performance in both zero-shot transfer and representation learning. Systematic evaluation is con- ducted spanning a wide variety of downstream tasks in- cluding zero-shot classification, out-of-domain classifica- tion, retrieval, visual representation learning, and textual representation learning, showcasing a consistent perfor- mance gain and validating the effectiveness of xCLIP. The code and pre-trained models will be publicly available at https://github.com/shallowtoil/xclip.	1. Introduction Language-image pre-training which simultaneously learns textual and visual representation from large-scale image-text pairs has revolutionized the field of representa- tion learning [25, 59], vision-language understanding [22], and text-to-image generation [61]. Compared to traditional visual models, language-instilled ones intrinsically inherit the capability of zero-shot or few-shot learning prominently demonstrated by large language models such as GPT-3 [12]. The precursor system, Contrastive Language- Image Pre- Training [59] ( CLIP ) that explicitly aligns the projected features of two modalities, has demonstrated surprising ca- pabilities of zero-shot, representation learning, and robust- ness, being applied to a wide range of fields [32,52,60,70]. …………512-dim …………512-dim CLIPImage EncoderProjection imagesLanguage EncoderProjection texts32768 -dim nCLIPImage EncoderProjection imagesLanguage EncoderProjection…… texts……32768 -dimattract distractFigure 1. Architecture comparison between CLIP and nCLIP. We take base -size encoders for an instance. CLIP discriminates instances within a batch using 512-dim projected embeddings. nCLIP projects each modality into a 32768-dim probability dis- tribution as the pseudo-label to supervise the prediction from the other modality. Darker blocks of nCLIP depict a higher response for cluster distribution, signifying the clusters to which text or im- age instances may belong. xCLIP is a multi-tasking framework with both CLIP and nCLIP. To learn from noisy web-crawled image-text pairs, CLIP adopts a formulation of the contrastive objective, where the image and text within a pair are considered as unique in- stances and are encouraged to be discriminated from all the other negative instances. However, web-crawled image-text pairs [65,69] are usually loosely correlated in the sense that one caption (image) can match reasonably with multiple im- ages (captions) besides the ground-truth one, as statistically shown in [74]. Hence, it is inaccurate and data-inefficient for representation learning to neglect other sensible matches and overlook the semantics hidden inside the textual de- scription. This is also solidified by the undesirable dis- criminative ability or transferring performance of the visual encoder pre-trained under the contrastive objective, as sug- gested in [78]. Mining semantics from plentiful concepts appearing in captions, therefore, solicits further exploration beyond the vanilla contrastive objective. Previously, some works resort to mining nearest neigh- bor positive samples via intra-modality similarity [50, 74] but require extra storage for auxiliary modules, i.e., the teacher network [74] or the memory bank [50]. Other works conduct multi-tasking with image-label supervision [57,78, 80], transforming the captions into tag format or the im- age tag into the captioning format for unified contrastive This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 11028 learning. Notwithstanding, such practice is still unable to account for the loose assignment within the image-text cap- tioning pairs. Intuitively, a caption for an image usually identifies existing objects within the image, which can be captured by a probabilistic distribution estimating how the text is assigned to one or multiple object clusters [13, 14]. Such estimation depicts the semantic meanings for its con- tained visual contents, and thus can be leveraged as a pseudo-label to guide the learning of the visual encoder. Inspired by the superiority of visual self-supervised learning (SSL) models pre-trained with the non-contrastive objective [14, 15], we explore whether the non-contrastive objective can be used across modalities for pre-training language-image models, and whether nice properties dis- played on visual SSL models can be inherited. To this end, we follow the same setup as CLIP except for the objective, and study non-Contrastive Language- Image Pre-Training (nCLIP ). For an image-text pair, we use the estimation of the textual (visual) distribution as the target to supervise the visual (textual) distribution, measured by cross-entropy. Additional regularizers are applied to avoid trivial solutions. Schematic comparison between CLIP and nCLIP is shown in Fig. 1. Theoretically, such formulation takes one modal- ity and the cluster centroid that the other modality belongs to as the positive samples to contrast [14], as opposed to direct features of two modalities as in contrastive learning, such that a single image is tolerated to be aligned with mul- tiple captions and concepts. Based on the above formula- tion, we conduct a systematic study in terms of zero-shot transfer and representation learning. We empirically ob- serve that, while nCLIP demonstrates desirable represen- tation learning capability for both modalities, it underper- forms prominently in zero-shot classification and retrieval tasks where models pre-trained with negative samples natu- rally prevail. Seeking for unique strengths of both worlds, we further perform multi-tasking of CLIP and nCLIP, short as xCLIP , and seek synergy between two distinct pre-training objec- tives. With the bearable add-on of computational resources (e.g.,∼27% memory-wise and ∼30% time-wise), xCLIP achieves consistent performance gain compared to CLIP on a wide range of tasks spanning from zero-shot classification, retrieval, linear probing, and fine-tuning, etc. An extensive study with different pre-training datasets, evaluation met- rics, and optimization configurations is conducted to val- idate xCLIP’s effectiveness in mining semantics and im- proving data efficiency. Particularly, the base-size model pre-trained using xCLIP under 35-million publicly avail- able image-text pairs achieves a performance gain of 3.3% and1.5% on an average of 27 classification tasks [59], in terms of zero-shot classification and linear probing accu- racy, respectively. Performance gain is also valid in terms of zero-shot retrieval, semi-supervised learning, and fine-tuning, with 3.7points of R@1 on Flickr30K [58], 1.7% accuracy on 1% subset of ImageNet [24], and 0.3% accu- racy on ImageNet, respectively.
Zhao_Exploring_Incompatible_Knowledge_Transfer_in_Few-Shot_Image_Generation_CVPR_2023	Abstract Few-shot image generation (FSIG) learns to generate di- verse and high-fidelity images from a target domain using a few ( e.g., 10) reference samples. Existing FSIG methods se- lect, preserve and transfer prior knowledge from a source generator (pretrained on a related domain) to learn the tar- get generator. In this work, we investigate an underexplored issue in FSIG, dubbed as incompatible knowledge transfer, which would significantly degrade the realisticness of syn- thetic samples. Empirical observations show that the issue stems from the least significant filters from the source gen- erator. To this end, we propose knowledge truncation to mitigate this issue in FSIG, which is a complementary op- eration to knowledge preservation and is implemented by a lightweight pruning-based method. Extensive experiments show that knowledge truncation is simple and effective, con- sistently achieving state-of-the-art performance, includingchallenging setups where the source and target domains are more distant. Project Page: yunqing-me.github.io/RICK.	1. Introduction Over recent years, deep generative models [14,15,21,30, 52, 78] have made tremendous progress, enabling many in- triguing tasks such as image generation [5, 27–29], image editing [45, 63, 68, 80], and data augmentation [6, 12, 59]. In spite of their remarkable success, most research on gen- erative models has been focusing on setups with sizeable training datasets [24, 26, 59, 60, 69, 74], limiting its applica- tions in many domains where data collection is difficult or expensive [12, 13, 46] ( e.g., medicine). To address such problems, FSIG has been proposed re- cently [41, 50], which learns to generate images with ex- tremely few reference samples ( e.g., 10 samples) from a tar- get domain. In these regimes, learning a generator to cap- ture the underlying target distribution is an undetermined This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 7380 problem that requires some prior knowledge. The majority of existing state-of-the-art (SOTA) FSIG methods rely on transfer learning approaches [2, 23, 65, 71] to exploit prior knowledge ( e.g., a source generator) learned from abundant data of a different but related source domain, and then trans- fer suitable source knowledge to learn the target generator with fine-tuning [9, 26, 41, 48–50, 64, 65, 74, 75, 77]. Differ- ent techniques have been proposed to effectively preserve useful source knowledge, such as freezing [48], regulariza- tion [41, 50, 77] and modulation [75] (details in Sec. 2). Incompatible knowledge transfer. Despite the impres- sive improvement achieved by different knowledge preser- vation approaches [41,48,50,75,77], in this work, we argue that preventing incompatible knowledge transfer is equally crucial. This is revealed through a carefully designed in- vestigation, where such incompatible knowledge transfer is manifested in the presence of unexpected semantic fea- tures. These features are inconsistent with the target do- main, thereby degrading the realisticness of synthetic sam- ples. As illustrated in Figure 1, trees and buildings are in- compatible with the domain of Sailboat (as can be observed by inspecting the 10 reference samples). However, they appear in the synthetic images when applying the existing SOTA methods [41, 75] with a source generator trained on Church . This shows that the existing methods cannot effec- tively prevent the transfer of incompatible knowledge. Knowledge truncation. Based on our observations, we propose Removing In-Compatible Knowledge (RICK), a lightweight filter-pruning based method to remove filters that encode incompatible knowledge ( i.e., filters with least estimated importance for adaptation) during FSIG adapta- tion. While filter pruning has been applied extensively to achieve compact deep networks with reduced computation [19,57,66], its application to prevent transfer of incompati- ble knowledge is underexplored. We note that our proposed knowledge truncation and pruning of incompatible filters are orthogonal and complementary with existing knowledge preservation methods in FSIG. In this way, our method ef- fectively removes the incompatible knowledge compared to prior works, and achieves noticeably improved quality ( e.g., FID [20]) of generated images. Our contributions can be summarized as follows: ‚We explore the incompatible knowledge transfer for FSIG, reveal that SOTA methods fail in handling this issue, investigate the underlying causation, and disclose the inade- quacy of fine-tuning in removal of incompatible knowledge. ‚We propose knowledge truncation to alleviate incom- patible knowledge transfer, and realize it with a lightweight filter-pruning based method. ‚Extensive experiments show that our method effec- tively removes incompatible knowledge and consistently improves the generative quality, including challenging se- tups where source and target domains are dissimilar.
Zhao_Towards_Better_Stability_and_Adaptability_Improve_Online_Self-Training_for_Model_CVPR_2023	Abstract Unsupervised domain adaptation (UDA) in semantic segmentation transfers the knowledge of the source domain to the target one to improve the adaptability of the seg- mentation model in the target domain. The need to access labeled source data makes UDA unable to handle adap- tation scenarios involving privacy, property rights protec- tion, and conﬁdentiality. In this paper, we focus on unsu- pervised model adaptation (UMA), also called source-free domain adaptation, which adapts a source-trained model to the target domain without accessing source data. We ﬁnd that the online self-training method has the potential to be deployed in UMA, but the lack of source domain loss will greatly weaken the stability and adaptability of the method. We analyze two reasons for the degradation of online self- training, i.e. inopportune updates of the teacher model and biased knowledge from the source-trained model. Based on this, we propose a dynamic teacher update mechanism and a training-consistency based resampling strategy to im- prove the stability and adaptability of online self-training. On multiple model adaptation benchmarks, our method ob- tains new state-of-the-art performance, which is compara- ble or even better than state-of-the-art UDA methods. The code is available at https://github.com/DZhaoXd/DT-ST.	1. Introduction Unsupervised Domain Adaptation (UDA) has received extensive attention on semantic segmentation tasks [49, 59, 60, 63], which transfers the knowledge in the source do- mains ( e.g. synthetic scene) to the target ones ( e.g. real scene). UDA in semantic segmentation aims to alleviate the dependence of deep neural network-based models on dense annotations [18,46,61] and improve their generaliza- This work is supported by the National Natural Science Foundation of China(No.62271377, No.62201407), the Key Research and Development Program of Shannxi (No.2021ZDLGY01-06, No.2022ZDLGY01-12), the National Key R &D Program of China (No. 2021ZD0110404), the China Postdoctoral Science Foundation (No. 2022M722496), the Foreign Schol- ars in University Research and Teaching Program’s 111 Project (B07048). Figure 1. Under the unsupervised model adaptation (UMA) set- ting, the mIoU score (%) of different methods on the validation set throughout the training in GTA5 !Cityscapes adaptation task. The dashed line represents the self-training UDA methods, and the solid line represents the MDA methods. tion ability to target domains [5, 12, 15]. However, in pro- prietary, privacy, or proﬁt-related concerns, source domain data is often unavailable, which presents new challenges for UDA [9, 27, 55]. To this end, the setting of Unsupervised Model Adaptation (UMA) is proposed [6,9,21,30,35], aim- ing to adapt the source-trained model to the unlabeled target domain without using source domain data. In UMA, the knowledge in the source-trained model becomes the only available supervision signal, making self-training on pseudo-labels the mainstream in the ﬁeld. Most existing UMA methods [26, 36, 57] adopt ofﬂine self- training methods, which iteratively updates the pseudo- labels and retrains the models. Although some improve- ments have been made, iterative self-training requires ex- pert intervention [1,59], as ill-suited rounds and termination often make it under-adapted. The recently proposed online self-training (ONST) methods [1, 31, 59] in UDA avoid the iterative training by online co-evolving pseudo labels, showing great potential. Then can ONST be applied to UMA scenarios without ac- cessing source data? We deploy the state-of-the-art ONST This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 11733 methods ProDA [1], SAC [59] and CPST [31] to UMA and draw the mIoU score curve on the validation set dur- ing training, as shown in Fig. 1. These ONST methods (dashed line) achieve more competitive performance than existing UMA methods (solid line). Nevertheless, taking a closer look at the curves in Fig. 1, these ONST methods present different degrees of degradation and unstable adap- tation process. Besides, their best performance in UMA de- creased by 4%5%mIoU scores on average than in UDA (See Table 1 and 2 in detail). Consequently, we conclude that existing ONST methods suffer from impaired stability and adaptability when applied to UMA. This paper is committed to improving the stability and adaptability of ONST methods in UMA. To begin with, we explore two reasons for the poor stability and adaptabil- ity of ONST in UMA. (1) The inopportune update of the teacher model causes the failure of co-evolution because the teacher model will continuously aggregate unevolved stu- dents. Concretely, as the teacher becomes the only supervi- sor in UMA, rapid updating will make the student lose the direction of evolution, and slow updating will make the stu- dent overﬁt the historical supervision, all of which leads to humble beneﬁts of teachers’ updating. (2) The bias towards minority categories in the source-trained model results in insufﬁcient adaptation to those minorities as the bias is eas- ily ampliﬁed in ONST, even with heuristic [1] or prototype thresholding [59] being set. Next, we present the explored solutions. For (1), we ﬁnd that the student model’s performance on historical samples during evolution can feedback on whether the student has evolved. Consequently, we propose a Dynamic Teacher Update (DTU) mechanism. DTU explores two feedback signals by information entropy [13] and soft neighborhood density [45], which can assess the evolutionary state of stu- dents. DTU then dynamically controls the update inter- val of the teacher model according to the students’ feed- back to aggregate more evolved students. For (2), we ﬁnd that resampling minority categories can effectively alleviate the bias towards minorities in UMA. However, most exist- ing resampling strategies [10, 11, 20, 50] rely on the source data and cannot apply in UMA. To this end, we propose a Training-Consistency based Resampling (TCR) strategy. TCR adaptively estimates the biased categories from the being-adapted model and selects reliable samples in biased categories as resampling candidates. Through these efforts, our method greatly improves the stability and adaptability of ONST in UMA, as shown in Fig. 1 (red solid line). We refer our method to DT-ST , asDTU and TCR play critical parts in online Self-Training under UMA. Sufﬁcient experiments show that DT-ST further exploits the potential of online self-training in UMA, towards bet- ter stability and adaptability. Moreover, DT-ST obtains new state-of-the-art performance on different UMA benchmarksand achieves comparable or even better performance than advanced UDA methods.
Zhao_DiffSwap_High-Fidelity_and_Controllable_Face_Swapping_via_3D-Aware_Masked_Diffusion_CVPR_2023	Abstract In this paper, we propose DiffSwap, a diffusion model based framework for high-fidelity and controllable face swapping. Unlike previous work that relies on carefully de- signed network architectures and loss functions to fuse the information from the source and target faces, we reformu- late the face swapping as a conditional inpainting task, per- formed by a powerful diffusion model guided by the desired face attributes (e.g., identity and landmarks). An impor- tant issue that makes it nontrivial to apply diffusion mod- els to face swapping is that we cannot perform the time- consuming multi-step sampling to obtain the generated im- age during training. To overcome this, we propose a mid- point estimation method to efficiently recover a reasonable diffusion result of the swapped face with only 2 steps, which enables us to introduce identity constraints to improve the face swapping quality. Our framework enjoys several fa- vorable properties more appealing than prior arts: 1) Con- trollable. Our method is based on conditional masked diffu- sion on the latent space, where the mask and the conditions can be fully controlled and customized. 2) High-fidelity. The formulation of conditional inpainting can fully exploit the generative ability of diffusion models and can preserve the background of target images with minimal artifacts. 3) †Corresponding authorShape-preserving. The controllability of our method en- ables us to use 3D-aware landmarks as the condition during generation to preserve the shape of the source face. Ex- tensive experiments on both FF++ and FFHQ demonstrate that our method can achieve state-of-the-art face swapping results both qualitatively and quantitatively.	1. Introduction There has been growing interest in face swapping tech- nology from both vision and graphics communities [1, 2, 6, 21,36,39] because of its broad applications in creating digi- tal twins, making films, protecting privacy, etc. The goal of face swapping is to transfer the identity of the source face to a target image or a video frame while keeping the attributes (e.g., pose, expression, background) unchanged. There are two essential steps in realizing high-quality face swapping: encoding the identity information of the source face effectively and blending identity and attributes from different images seamlessly. Early work [4,24] on face swapping adopts 3D models [5] to represent the source face and directly replace the reconstructed faces in the target im- age based on 3D structural priors, leading to recognizable artifacts. The development of generative adversarial net- works (GAN) [12] provides a strong tool to generate photo- realistic face images. Many recent methods [2, 6, 21, 39] perform face swapping by extracting the identity feature This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 8568 from the source image and then injecting it into the gen- erative models powered by adversarial training. However, these methods tend to make minor modifications to the tar- get image, which may fail in totally transferring the identity information when the face shapes of the source and the tar- get face largely differ. Very recently, diffusion-based models (DM) [22, 25, 26] have exhibited high customizability for various conditions and impressive power in generating images with high res- olution and complex scenes. It is natural to ask: whether the strong generation ability of diffusion models can benefit face swapping? However, we find it is nontrivial to apply diffusion models to the task of face swapping. Since there is no ground-truth data pair for face swapping, face swapping models are usually trained in a weakly-supervised manner, where several losses about image fidelity, identity, and fa- cial attributes are imposed to guide the training. These su- pervisory signals can be easily added to GAN-based models but it is difficult for DMs. Different from previous genera- tive models like GANs [12, 16] and V AEs [14, 19], DMs learn a denoising autoencoder to gradually recover the data density step-by-step. Although the autoencoder can be ef- ficiently learned by performing score matching [15] at an arbitrary step during training, image generation using an al- ready trained DM requires executing the autoencoder se- quentially for a large number of steps (typically, 200 steps), which is computationally expensive. To tackle these challenges, we propose the first diffusion model based face swapping framework, which can produce high-fidelity results faces with high controllability. Figure 2 shows the overview of our method. Different from existing methods [1,6,21,36] that modify the target face to match the identity of the source face, we reformulate face swapping as a conditional inpainting task guided by the identity fea- ture and facial landmarks. Our diffusion model is learned to generate a face that shares the same identity as the source face and is spatially aligned with the target face. In order to introduce identity constraints during training, we propose a midpoint estimation method that can efficiently generate swapped faces with only 2 steps. Our framework is by de- sign highly controllable, where both the conditioned land- mark and the inpainting mask can be customized during in- ference. Thanks to this property, we propose the 3D-aware masked diffusion where we perform the inpainting inside the 3D-aware mask conditioned on the 3D-aware landmark that explicitly enforces the shape consistency between the source face and the swapped face. We conducted extensive experiments on FaceForen- sics++ [27] and FFHQ [16] to verify the effectiveness of our model both and quantitatively. On FF++ dataset, our method outperforms previous methods in both ID retrieval (98.54%) and FID (2.16), while achieving comparable re- sults on pose error and expression error. Qualitative resultsshow that our method can generate high-fidelity swapped faces that can better preserve the source face shape than the previous method. Besides, we also demonstrate the scala- bility and controllability of our method. Our model can be easily extended to higher-resolution such as 512×512with affordable extra computational costs and allows region- customizable face swapping by controlling the inpainting mask. Our results demonstrate that DiffSwap is a very promising face swapping framework that is distinct from the existing methods and enjoys high fidelity, controllabil- ity, and scalability.
Zhu_TryOnDiffusion_A_Tale_of_Two_UNets_CVPR_2023	Abstract Given two images depicting a person and a garment worn by another person, our goal is to generate a visu- alization of how the garment might look on the input per- son. A key challenge is to synthesize a photorealistic detail- preserving visualization of the garment, while warping the garment to accommodate a significant body pose and shape change across the subjects. Previous methods either fo- cus on garment detail preservation without effective pose 1Work done while author was an intern at Google.and shape variation, or allow try-on with the desired shape and pose but lack garment details. In this paper, we pro- pose a diffusion-based architecture that unifies two UNets (referred to as Parallel-UNet), which allows us to preserve garment details and warp the garment for significant pose and body change in a single network. The key ideas behind Parallel-UNet include: 1) garment is warped implicitly via a cross attention mechanism, 2) garment warp and person blend happen as part of a unified process as opposed to a se- quence of two separate tasks. Experimental results indicate that TryOnDiffusion achieves state-of-the-art performance This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 4606 both qualitatively and quantitatively.	1. Introduction Virtual apparel try-on aims to visualize how a garment might look on a person based on an image of the person and an image of the garment. Virtual try-on has the potential to enhance the online shopping experience, but most try-on methods only perform well when body pose and shape vari- ation is small. A key open problem is the non-rigid warping of a garment to fit a target body shape, while not introducing distortions in garment patterns and texture [5, 12, 41]. When pose or body shape vary significantly, garments need to warp in a way that wrinkles are created or flat- tened according to the new shape or occlusions. Related works [1,5,23] have been approaching the warping problem via first estimating pixel displacements, e.g., optical flow, followed by pixel warping, and postprocessing with percep- tual loss when blending with the target person. Fundamen- tally, however, the sequence of finding displacements, warp- ing, and blending often creates artifacts, since occluded parts and shape deformations are challenging to model ac- curately with pixel displacements. It is also challenging to remove those artifacts later in the blending stage even if it is done with a powerful generative model. As an alternative, TryOnGAN [24] showed how to warp without estimating displacements, via a conditional StyleGAN2 [21] network and optimizing in generated latent space. While the gener- ated results were of impressive quality, outputs often lose details especially for highly patterned garments due to the low representation power of the latent space. In this paper, we present TryOnDiffusion that can handle large occlusions, pose changes, and body shape changes, while preserving garment details at 1024×1024 resolution. TryOnDiffusion takes as input two images: a target person image, and an image of a garment worn by another person. It synthesizes as output the target person wearing the gar- ment. The garment might be partially occluded by body parts or other garments, and requires significant deforma- tion. Our method is trained on 4 Million image pairs. Each pair has the same person wearing the same garment but ap- pears in different poses. TryOnDiffusion is based on our novel architecture called Parallel-UNet consisting of two sub-UNets communicating through cross attentions [40]. Our two key design elements are implicit warping and combination of warp and blend (of target person and garment) in a single pass rather than in a sequential fashion. Implicit warping between the target per- son and the source garment is achieved via cross attention over their features at multiple pyramid levels which allows to establish long range correspondence. Long range corre- spondence performs well, especially under heavy occlusion and extreme pose differences. Furthermore, using the same network to perform warping and blending allows the twoprocesses to exchange information at the feature level rather than at the color pixel level which proves to be essential in perceptual loss and style loss [19, 29]. We demonstrate the performance of these design choices in Sec. 4. To generate high quality results at 1024×1024 resolu- tion, we follow Imagen [35] and create cascaded diffusion models. Specifically, Parallel-UNet based diffusion is used for128×128and256×256resolutions. The 256×256re- sult is then fed to a super-resolution diffusion network to create the final 1024×1024 image. In summary, the main contributions of our work are: 1) try-on synthesis at 1024×1024 resolution for a variety of complex body poses, allowing for diverse body shapes, while preserving garment details (including patterns, text, labels, etc.), 2) a novel architecture called Parallel-UNet, which can warp the garment implicitly with cross atten- tion, in addition to warping and blending in a single net- work pass. We evaluated TryOnDiffusion quantitatively and qualitatively, compared to recent state-of-the-art methods, and performed an extensive user study. The user study was done by 15non-experts, ranking more than 2K distinct ran- dom samples. The study showed that our results were cho- sen as the best 92.72% of the time compared to three recent state-of-the-art methods.
Zhou_Efficient_Second-Order_Plane_Adjustment_CVPR_2023	Abstract Planes are generally used in 3D reconstruction for depth sensors, such as RGB-D cameras and LiDARs. This paper focuses on the problem of estimating the optimal planes and sensor poses to minimize the point-to-plane distance. The resulting least-squares problem is referred to as plane ad- justment (PA) in the literature, which is the counterpart of bundle adjustment (BA) in visual reconstruction. Iterative methods are adopted to solve these least-squares problems. Typically, Newton’s method is rarely used for a large-scale least-squares problem, due to the high computational com- plexity of the Hessian matrix. Instead, methods using an ap- proximation of the Hessian matrix, such as the Levenberg- Marquardt (LM) method, are generally adopted. This pa- per adopts the Newton’s method to efficiently solve the PA problem. Specifically, given poses, the optimal plane have a close-form solution. Thus we can eliminate planes from the cost function, which significantly reduces the number of variables. Furthermore, as the optimal planes are func- tions of poses, this method actually ensures that the optimal planes for the current estimated poses can be obtained at each iteration, which benefits the convergence. The diffi- culty lies in how to efficiently compute the Hessian matrix and the gradient of the resulting cost. This paper provides an efficient solution. Empirical evaluation shows that our algorithm outperforms the state-of-the-art algorithms.	1. Introduction Planes ubiquitously exist in man-made environments, as demonstrated in Fig. 1. Thus they are generally used in simultaneous localization and mapping (SLAM) systems for depth sensors, such as RGB-D cameras [6, 9, 12–14] and LiDARs [16, 21, 22, 24, 26]. Just as bundle adjust- ment (BA) [3, 8, 20, 25] is important for visual reconstruc- tion [1,5,18,19], jointly optimizing planes and depth sensor poses, which is called plane adjustment (PA) [24,26], is crit- ical for 3D reconstruction using depth sensors. This paper focuses on efficiently solving the large-scale PA problem. The BA and PA problems both involve jointly optimiz- Figure 1. We use Gaussian noises to perturb the poses of dataset D in Fig. 3. The standard deviations for rotation and translation are3◦and0.3m, respectively. The resulting point cloud (a) is in a mess. Fig. (b) shows the result from our algorithm. Our algorithm can quickly align the planes, as shown in Fig. 5. ing 3D structures and sensor poses. As the two problems are similar, it is straightforward to apply the well-studied solu- tions for BA to PA, as done in [23, 26]. However, planes in PA can be eliminated, so that the cost function of the PA problem only depends on sensor poses, which significantly reduces the number of variables. This property provides a promising direction to efficiently solve the PA problem. However, it is difficult to compute the Hessian matrix and the gradient vector of the resulting cost. Although this prob- lem was studied in several previous works [10, 16], no effi- cient solution has been proposed. This paper seeks to solve this problem. The main contribution of this paper is an efficient PA solution using Newton’s method. We derive a closed-form solution for the Hessian matrix and the gradient vector for the PA problem whose computational complexity is inde- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 13113 pendent of the number of points on the planes. Our experi- mental results show that the proposed algorithm converges faster than the state-of-the-art algorithms.
Zhuang_Deep_Semi-Supervised_Metric_Learning_With_Mixed_Label_Propagation_CVPR_2023	Abstract Metric learning requires the identification of far-apart similar pairs and close dissimilar pairs during training, and this is difficult to achieve with unlabeled data because pairs are typically assumed to be similar if they are close. We present a novel metric learning method which circumvents this issue by identifying hard negative pairs as those which obtain dissimilar labels via label propagation (LP), when the edge linking the pair of data is removed in the affinity matrix. In so doing, the negative pairs can be identified despite their proximity, and we are able to utilize this infor- mation to significantly improve LP’s ability to identify far- apart positive pairs and close negative pairs. This results in a considerable improvement in semi-supervised metric learning performance as evidenced by recall, precision and Normalized Mutual Information (NMI) performance met- rics on Content-based Information Retrieval (CBIR) appli- cations.	1. Introduction Image data have proliferated owing to ubiquitous cam- era ownership, widespread internet connectivity, and broad availability of software applications. The design of efficient CBIR systems is imperative in managing information con- tent. Traditional image retrieval systems use textual annota- tions, but such an approach require manual labor which may be not be cost-effective. On the other hand, CBIR retrieves relevant content from an image query, alleviating the need for human annotation. Metric learning learns a transformation mapping similar data closer than dissimilar ones. This makes it easier to cluster data [37] and also to shortlist similar data by their proximity to the query. As such, metric learning algorithms are at the cornerstone of most state-of-the-art CBIR designs [23]. As the amount of data handled in CBIR systems is very large, there is interest in utilizing the vast quantities * Work done while a student at UIUC.of readily available unlabeled data to improve the metric learning training process. Recent advances have asserted the importance of retain- ing within-class variance in training generalizable represen- tations [27], particularly because the information encapsu- lated in representations which are irrelevant in discriminat- ing between training labels may be important in discrim- inating between unseen test classes. Since close positive pairs results in little loss, there is a need to identify far- apart positive pairs for training. The paper [43] uses the method from [9] to propagate labels along manifolds in order to identify such points, by assuming data which are nearest neighbor pairs but not mutual nearest neighbors as close dissimilar pairs. However this method does not uti- lize label information and may not accurately identify hard negatives [3, 17, 39, 40], which are pairs of data which are close but have different labels. Label propagation (LP) us- ing mis-identified edges would lead to inaccuracies in the model trained. This paper proposes a novel method to identify hard neg- ative pairs in a semi-supervised setting. Specifically, given a few labeled points from each class and an abundance of un- labeled data, we propose to identify hard negative pairs as those which obtain dissimilar labels via LP when the edge linking the two elements of the pair is removed in the affin- ity matrix. We obtain the dissimilarity weights of edges under this assumption quickly and efficiently, without the costly use of repeated LP to calculate these weights. We use this negative edge information in a novel mixed LP algorithm which is able to utilize both positive and nega- tive edge information. Specifically, the method encourages data linked by positive edges to have the same pseudolabel and data linked by negative edges to have different pseu- dolabels. Like LP, our mixed LP optimization problem can be solved with conjugate gradient (CG), allowing it to scale to large datasets. As our obtained pseudolabels capture information on far- apart positive pairs and close negative pairs, they yield a significant improvement in semi-supervised metric learn- ing. We showcase this in a CBIR setting under recall, pre- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 3429 cision and Normalized Mutual Information (NMI) perfor- mance metrics. This shows that our method is able to more effectively rank the database in relation to the query and re- turn relevant articles near the top of this list.
Zhou_HyperMatch_Noise-Tolerant_Semi-Supervised_Learning_via_Relaxed_Contrastive_Constraint_CVPR_2023	Abstract Recent developments of the application of Contrastive Learning in Semi-Supervised Learning (SSL) have demon-strated signiﬁcant advancements, as a result of its ex-ceptional ability to learn class-aware cluster representa-tions and the full exploitation of massive unlabeled data.However , mismatched instance pairs caused by inaccuratepseudo labels would assign an unlabeled instance to theincorrect class in feature space, hence exacerbating SSL’srenowned conﬁrmation bias. To address this issue, we intro-duced a novel SSL approach, HyperMatch, which is a plug-in to several SSL designs enabling noise-tolerant utilizationof unlabeled data. In particular , conﬁdence predictions arecombined with semantic similarities to generate a more ob-jective class distribution, followed by a Gaussian MixtureModel to divide pseudo labels into a ’conﬁdent’ and a ’lessconﬁdent’ subset. Then, we introduce Relaxed ContrastiveLoss by assigning the ’less-conﬁdent’ samples to a hyper-class, i.e. the union of top- Knearest classes, which effec- tively regularizes the interference of incorrect pseudo la-bels and even increases the probability of pulling a ’less conﬁdent’ sample close to its true class. Experiments and in-depth studies demonstrate that HyperMatch delivers re-markable state-of-the-art performance, outperforming Fix-Match on CIF AR100 with 400 and 2500 labeled samples by11.86% and4.88% , respectively.	1. Introduction Semi-supervised learning (SSL) [ 2,3,5,20,21,26,29, 38,39,42] has become a promising solution for leverag- ing unlabeled data to save the expensive annotation costand simultaneously improve model performance, especially in applications where large amounts of annotated data are required to obtain a model with high performance [ 4,18]. Modern SSL algorithms generally fall into two categories:the Pseudo Label -based [ 22,30,38] focuses on generating ∗Equal contribution.†Corresponding author.:UV1'II[XGI_UL 6YK[JU2GHKRY? 6YK[JU2GHKRHGYKJ )RGYY'YYOMTSKTZH._VKX)RGYY'YYOMTSKTZ? G6[RR 6[YN I Figure 1. Illustration of our ideas. (a)Pseudo-label-based assign- ment: the instance is pulled close to the wrong pseudo label class and pushed away from ground-truth class (green pentagon). (b) Hyper-class assignment: the instance is assigned to its hyper-class(the union of top- Knearest classes), which includes the ground truth class. (c)Top-Kaccuracy for clean and noisy pseudo labels (divided by our Gaussian Mixture Model) in CIFAR100@400 ex-periment. As Kgrows, noisy labels beneﬁt more than clean labels. reliable pseudo labels for unlabeled data, whereas the Con- sistency Regularization -based [ 2,3,20,29] constrains the model to make consistent predictions on perturbed samples. Recently, a prominent advance is the combination of Contrastive Learning [ 6,8,11,13] with SSL techniques [19,21,24,25,39,42], which sets the remarkable state-of- the-art performance. The naive Self-supervised ContrastiveLearning [ 6,13] in pre-training tasks pushes instance-level features away, thereby potentially driving samples withinthe same class apart, its class-agnostic nature has been proved to conﬂict with the class-clustering property ofSSL [ 24,39], hence most recent studies [ 21,24,39] turn to Class-Aware Contrastive Learning [ 16]. The general rou- tine is to ﬁrst assign each unlabeled instance to the top-1nearest class based on their pseudo labels, then two unla-beled instances from the same class are randomly selected to form a positive pair, followed by a class-aware con- trastive loss to compel instances from positive pairs to sharesimilar features while pushing away the representations ofdifferent classes. The precision of pseudo labels has a direct impact on the class assignment of the aforementioned methods. By con-ﬁdence threshold, unlabeled data can be roughly groupedinto ’conﬁdent’ and ’less conﬁdent’ [ 39]. For ’conﬁdent’ This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 24017 data that tends to yield accurate pseudo labels for class as- signment, contrastive loss could constrain the model to ac-quire better clustered feature representations for each class,hence facilitating SSL learning. But for ’less conﬁdent’ datawith a much higher probability to provide incorrect pseudo labels, mismatched instance pairs tends to be induced and the contrastive constraint will pull the features of differ-ent classes closer while pushing the features from the sameclass further apart, which will inevitably degrade the learn-ing. As shown in Fig. 1 (a) , the ’less conﬁdent’ instance could be drawn close to a false class ( i.e. the red circle) while being pushed away from the true class, representedby green pentagons. For convenience, we also refer to the two kinds of unlabeled data as ’clean’ and ’noisy’ data. To mitigate the detrimental effects of ’less conﬁdent’ (i.e., noisy) data, existing studies can be generally cat- egorized into two categories: (1) Discarding low-qualitypseudo labels [ 21,24] with a low threshold, leaving some unlabeled data unused; (2) Adopting re-weighting strate-gies [ 39,42] to lessen the effect of noisy labels. How- ever, these approaches continue to closely adhere to thethe paradigm of assigning noisy data to a single error-proneclass, therefore they can only reduce the errors to a limitedamount. Also, the accuracy of pseudo labels suffers fromconﬁrmation bias [ 1] in SSL, which is the accumulation of false predictions during training. The aforesaid interference of ’less conﬁdent’ data is caused by the inaccurate class assignment, it is more ef-fective to devise a class assignment approach that can resistthe distraction of wrong pseudo labels in order to mitigatetheir effects. In light of this, we propose to relax the conven-tional class assignment. Instead of assigning a noisy sam-ple to a single class, we relax the assignment by grouping itinto a ’hyper-class’, which is the union of top- K(K>1) nearest classes. As depicted in Fig. 1 (b) , the chance of the ground-truth class slipping into the hyper-class is dra-matically enhanced by implementing the relaxation (as K steadily grows). It’s also worth noting that the marginal gain brought by relaxation for noisy data is signiﬁcantly greaterthan that for clean data, as the top-1 pseudo label accuracyof clean data is already adequate. This suggests that therelaxation is more suitable for applying on noisy data. In conjunction with the hyper-class assignment, a Re- laxed Contrastive Loss is intended to restrain the featureof noisy samples being close to their corresponding hyper-class while increasing the distantce from the remainingclasses. The simple yet effective relaxing has two bene-ﬁts: (1) the likelihood of ’less conﬁdent’ data being pushedaway from its ground truth class can be lowered, and (2) thelikelihood of data being pulled close to its ground truth classcan be successfully raised. As seen in Fig. 1 (b) , the ground truth class for the noisy unlabeled instance falls within thehyper-class, thus its feature will no longer be driven awayfrom the actual class, but rather drawn close to it. In order to manage the effective exploitation of bother clean and noisy unlabeled data, we proposed HyperMatch . First, predicted class probabilities are integrated with se- mantic similarities to produce unbiased per-sample class distributions. Next, a Gaussian Mixture Model (GMM)model is ﬁtted on the calibrated distribution to separateclean unlabeled data from the noisy ones. The commonclass-aware contrastive loss is applied to clean data to con-strain their features to approach the corresponding class.For noisy unlabeled data, a Relaxed Contrastive Loss iscarefully developed to drive the noisy unlabeled data fallinginto their corresponding hyper-class. In summary, we con-tribute in three ways: • We propose an enhanced contrastive learning method, HyperMatch, to handle the effective separation and ex-ploitation of both clean and noisy pseudo labels forlearning better-clustered feature representations. It is a plug-in that can be used to various SSL architecturesto increase resilience while utilizing noisy data. • Unlike previous studies that assign a ’less conﬁdent’ sample to an error-prone class, we relax the assignmentby categorizing the noisy sample into a hyper-class (aunion of top- Knearest classes), followed by the pro- posed Relaxed Contrastive Loss, which is effective atmitigating the problematic conﬁrmation bias. • With thorough experiments on SSL benchmarks, our HyperMatch demonstrates competitive performanceand establishes the new state-of-the-art on multiplebenchmarks. In-depth investigation reveals its effec-tiveness in handling the noisy pseudo labels.
Zhao_Comprehensive_and_Delicate_An_Efficient_Transformer_for_Image_Restoration_CVPR_2023	Abstract Vision Transformers have shown promising performance in image restoration, which usually conduct window- or channel-based attention to avoid intensive computations. Although the promising performance has been achieved, they go against the biggest success factor of Transformers to a certain extent by capturing the local instead of global de- pendency among pixels. In this paper, we propose a novel efficient image restoration Transformer that first captures the superpixel-wise global dependency, and then transfers it into each pixel. Such a coarse-to-fine paradigm is imple- mented through two neural blocks, i.e., condensed attention neural block (CA) and dual adaptive neural block (DA). In brief, CA employs feature aggregation, attention computa- tion, and feature recovery to efficiently capture the global dependency at the superpixel level. To embrace the pixel- wise global dependency, DA takes a novel dual-way struc- ture to adaptively encapsulate the globality from superpix- els into pixels. Thanks to the two neural blocks, our method achieves comparable performance while taking only ∼6% FLOPs compared with SwinIR.	1. Introduction Image restoration aims to recover the high-quality im- age from its degraded version, and huge success has been achieved by plentiful methods in the past years [22, 30, 35, 54–56, 60]. In the era of deep learning, Convolutional Neu- ral Networks (CNNs) have shown promising performance in image restoration [25, 52, 67] thanks to the inductive bi- ases of weight sharing and spatial locality [12]. However, although a number of studies have shown the effectiveness of CNNs, it has suffered from the following limitations [12], i.e., i) non-dynamic weights of CNNs limit the model ca- pacity of instance adaption, and ii) the sparse connections of CNNs limit the capture of global dependency. *indicates equal contribution. †Corresponding author. Superpixel-wise GlobalitySpatial LocalityChannel LocalityShifted Spatial Locality Sequentially Recover&DistributePixel-wise Globality Local Dependency Capture Global Dependency Capture (Ours)AggregateFigure 1. Illustration of the dependency capture in existing vision Transformers and ours. The red boxes refer to the dependency capture range of a given pixel marked by a red point. We gener- ally summarize the dependency capture in existing vision Trans- formers as the three ways above the dashed line. Obviously, they could only capture the dependency in a local range. In contrast, our method could obtain a pixel-wise global dependency through superpixel-wise dependency computation and distribution. To overcome these limitations, some solutions [51, 60, 67, 68] have been specifically established, of which Transformer-based methods [3, 5, 28, 47, 53] have achieved huge success thanks to their high capacities of dynamic weighting and global dependency capturing. Towards the image-restoration-specific Transformers, the biggest road- block might be the unacceptable cost caused by the global attention computation. Therefore, some efficient attentions have been proposed to trade-off the efficiency and depen- dency range, e.g., local window attention [49], shifted win- dow attention [22], and channel attention [56]. Although the promising performance has been achieved, these Trans- formers have still suffered from the following limitations. First, the computation costs are still very high, thus limit- ing their applications in the mobile scenarios. Second, the attention mechanisms could only capture the dependency from a given range, and thus such a locality might not fully explore the potential of Transformer. This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 14122 In practice, it is daunting to develop an efficient image restoration Transformer while directly capturing the global dependency, since it necessarily introduces intensive com- putations, which goes against efficiency. Different from ex- isting studies focusing on efficient attention mechanisms, this paper resolves the contradiction from a novel perspec- tive. To be specific, our basic idea is to adaptively aggregate the features at pixel level into a lower dimensional space of superpixel to remove the redundancy in channel [13] and space [50] domains. Through the feature aggregation, the dimension is remarkably reduced, and thus the attention could be computed in a global way with acceptable com- putation costs. After that, the feature recovery is performed to restore the feature distribution in channel and space do- mains. Although the above paradigm shows a feasible so- lution, it is far from the final goal since the obtained de- pendency actually works at the superpixel level. Hence, we need to transfer such a superpixel-wise global dependency to a pixel-wise global dependency. As a result, pixel-wise restoration can depend on the global information from su- perpixels. Based on the above motivations, we propose an effi- cient Transformer for COmprehensive and DElicate image restoration (CODE). To be specific, CODE consists of a condensed attention neural block (CA) and a dual adap- tive neural block (DA). In brief, CA captures the global dependency at the superpixel level with acceptable compu- tations thanks to the aforementioned paradigm. To obtain the pixel-wise global dependency, DA extracts the global- ity from the CA output and then dynamically encapsulates it into each pixel. Thanks to the two complementary neural blocks, CODE could capture the pixel-wise global depen- dency, while embracing high computational efficiency. The contributions and novelty of this work could be sum- marized as below. • Unlike existing efficient Transformers only capture the pixel-wise local dependency, our solution could obtain the pixel-wise global dependency through superpixel- wise dependency computation and transformation. • The proposed image restoration Transformer (CODE) consists of two neural blocks. In brief, CA employs feature aggregation, attention computation, and fea- ture recovery to efficiently capture the superpixel-wise global dependency. To obtain pixel-wise global depen- dency, DA takes a novel dual-way structure and a dy- namic weighting fashion to distribute the superpixel- wise globality into each pixel. • Extensive experiments are conducted on four image restoration tasks to demonstrate the efficiency and ef- fectiveness of CODE, e.g., it achieves comparable per- formance while taking only ∼6% FLOPs compared with SwinIR [22].
Zhou_BEVDC_Birds-Eye_View_Assisted_Training_for_Depth_Completion_CVPR_2023	Abstract Depth completion plays a crucial role in autonomous driving, in which cameras and LiDARs are two complemen- tary sensors. Recent approaches attempt to exploit spatial geometric constraints hidden in LiDARs to enhance image- guided depth completion. However, only low efficiency and poor generalization can be achieved. In this paper, we pro- pose BEV@DC , a more efficient and powerful multi-modal training scheme, to boost the performance of image-guided depth completion. In practice, the proposed BEV@DC model comprehensively takes advantage of LiDARs with rich geometric details in training, employing an enhanced depth completion manner in inference, which takes only im- ages (RGB and depth) as input. Specifically, the geometric- aware LiDAR features are projected onto a unified BEV space, combining with RGB features to perform BEV com- pletion. By equipping a newly proposed point-voxel spatial propagation network (PV-SPN), this auxiliary branch intro- duces strong guidance to the original image branches via 3D dense supervision and feature consistency. As a result, our baseline model demonstrates significant improvements with the sole image inputs. Concretely, it achieves state-of- the-art on several benchmarks, e.g., ranking Top-1 on the challenging KITTI depth completion benchmark.	1. Introduction Dense depth estimation plays an essential role in various 3D vision tasks and self-driving applications, e.g., 3D ob- ject detection and tracking, simultaneous localization and mapping (SLAM), and structure-from-motion (SFM) [14, 17,19,33,37]. With the aid of outdoor LiDAR sensors or in- door RGBD cameras, 3D vision applications acquire depth maps for further industrial usage. However, the depth sen- sors cannot provide dense pixel-wise depth maps since their output is sparse and has numerous blank regions, especially *Corresponding author Training Phase Inference Phase Single-modal Model Fusion-based Model Camera-based Model (c) OursFusion-based ModelFusion-based Model BEV assisted training(b) Fusion-based Training and inferenceVarious 3D inputRGB and Depth Single-modal Model Camera-based Model Single-modal Model Camera-based Model (a) Camera-based BEV and VoxelsFigure 1. BEV assisted training. (a) Previous camera-based methods that take RGB and depth input. (b) Previous fusion-based methods introduce extra inputs and computation in both training and inference. (c) Our method takes additional LiDAR as input for assisted training. Only the 2D inputs are used during the infer- ence, which reduces the computational burden. in outdoor scenes. Therefore, it is necessary to fill the void areas of the depth maps in practice. Recent depth completion methods [4, 12, 21, 47] lever- age the RGB information as guidance since the RGB im- ages contain scene structures, e.g., textures, and monoc- ular features, e.g., vanishing points, to provide the cues for the missing pixels. However, the camera-based meth- ods apply the 2D convolution on the irregularly distributed depth values, resulting in an implicit yet ineffective ex- ploration of underlying 3D geometry, i.e., over-smooth at the boundary of objects. Considering the deployment of cameras and LiDAR in commercial cars and the recent trend of cross-modal learning in the vision community, some methods [2, 3, 28, 41] introduce explicit 3D represen- tations, i.e., LiDAR point clouds generated by sparse depth, This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 9233 to complement 2D appearance features with 3D structured priors. Despite the improvements, the fusion-based ap- proaches still have the following issues: 1)The 3D fea- ture extraction and fusion are not efficacious, especially the critical spatial correlations between a depth point and its neighbors, which significantly affects the completion performance. 2)Fusion-based methods are computation- intensive while processing sparse depths, RGB images, and additional 3D input such as LiDAR information, either oc- cupying more memory storage or consuming more time in inference, which hinders real-time applications. To address the above issues, we seek to boost image- guided depth completion performance by exploiting 3D representations via a more efficient and effective cross- representation training scheme. In training, we design an auxiliary LiDAR branch consisting of LiDAR encoder, cross-representation BEV decoder (CRBD) and point-voxel spatial propagation network (PV-SPN). Initially, we prepro- cess each LiDAR scan with the assigned voxel cells to al- leviate the irregularity and sparseness and then extract its multi-scale features. After that, these features will be pro- jected onto a unified BEV space. The following CRBD uti- lizes the above multi-scale BEV features and the ones from the camera branch to perform BEV fusion and completion. After that, the BEV completion is interpolated into the 3D space, and a point-voxel spatial propagation network is pro- posed to query the nearest neighbors for each coarse voxel and performs feature aggregation on all the adjacent points from LiDAR, refining the 3D geometric shapes. Moreover, to tackle the extra computational burden from the LiDAR branch, this plug-and-play component is only exploited in the training phase, enhancing the original camera branch through feature consistency and end-to-end backpropaga- tion. Consequently, the trained model is independent of ad- ditional LiDAR inputs during the inference. Compared with previous fusion-based methods, our pro- posed framework has the following advantages: 1) Gener- ality : Our plug-and-play solution can be incorporated into several camera-based depth completion models; 2) Flexi- bility : The processing module for LiDAR representations only exists during training and is discarded in inference, as shown in Fig. 1(c), compared with previous camera-based models (a) and fusion-based models (b). There is no addi- tional computational burden in the deployment. 3) Effec- tiveness : It significantly boosts the performance upon the baseline approach, achieving state-of-the-art results on sev- eral benchmarks. To sum up, the main contributions are summarized as follows: -Bird’s- EyeView Assisted Training for Depth Completion (BEV@DC) is proposed, which as- sists camera-based depth completion with LiDAR representation during the training phase. - Cross-representation BEV decoder (CRBD) and point-voxel spatial propagation network (PV-SPN) are pro- posed to gain fine-grained 3D geometric shapes and provide strong guidance to the RGB branch. - Our solution achieves state-of-the-art on both outdoor KITTI depth completion benchmark and indoor NYU Depth v2 dataset.
Zhou_Adaptive_Sparse_Pairwise_Loss_for_Object_Re-Identification_CVPR_2023	Abstract Object re-identification (ReID) aims to find instances with the same identity as the given probe from a large gallery. Pairwise losses play an important role in training a strong ReID network. Existing pairwise losses densely exploit each instance as an anchor and sample its triplets in a mini-batch. This dense sampling mechanism inevitably introduces positive pairs that share few visual similarities, which can be harmful to the training. To address this prob- lem, we propose a novel loss paradigm termed Sparse Pair- wise (SP) loss that only leverages few appropriate pairs for each class in a mini-batch, and empirically demon- strate that it is sufficient for the ReID tasks. Based on the proposed loss framework, we propose an adaptive pos- itive mining strategy that can dynamically adapt to diverse intra-class variations. Extensive experiments show that SP loss and its adaptive variant AdaSP loss outperform other pairwise losses, and achieve state-of-the-art performance across several ReID benchmarks. Code is available at https://github.com/Astaxanthin/AdaSP.	1. Introduction Object re-identification (ReID) is one of the most impor- tant vision tasks in visual surveillance. It aims at associat- ing person/vehicle images with the same identity captured by different cameras in diverse scenarios. Recently, with the prosperity of deep neural networks in computer vision community, ReID tasks have rapidly progressed towards more sophisticated feature extractors [8, 37, 40, 44, 60, 61] and more elaborate losses [43, 45, 57, 58]. Benefiting from quantifying the semantic similarity/distance of two images, metric losses are widely employed with the identity loss to improve ranking precision in the ReID task [35]. Generally, metric losses serve for metric learning that aims to map raw signals into a low-dimensional embedding space where instances of intra-class are clustered and that of inter-class are separated. The pairwise framework, such †Corresponding author. Dense paradigm Sparse paradigmTriplet loss Circle loss SP lossFigure 1. Difference between dense pairwise losses and our sparse pairwise (SP) loss. For simplicity, only two classes, marked by circles and triangles, are displayed. The red circles represent an- chors. Dense pairwise losses, such as triplet loss [19] and circle loss [45], treat each instance as an anchor and mine its positive and negative pairs to construct a loss item. SP loss treats each class as a unit and separately excavates the most informative positive and negative pairs without using anchors. as triplet loss [19] and circle loss [45], employs anchors to pull their positive pairs while pushing negatives apart. Most existing pairwise losses [14, 19, 38, 45] devote to exploiting all instances in a mini-batch, and densely anchoring each of them (which we term dense pairwise method) to sam- ple its triplets. Although some hard sample mining meth- ods [7, 45, 50, 51] are developed to accelerate convergence or improve performance, these losses are still computed in a dense manner ( i.e. top rows in Fig. 1). This dense design in- evitably introduces harmful positive pairs [1, 43] that share few visual similarities and likely lead to bad local minima of optimization [54] for metric learning, due to large intra- class variations (Fig. 2a) caused by illumination changes, occlusion, different viewpoints, etc. In this work, we anticipate that it is not necessary to em- ploy all instances within a mini-batch since most of them This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 19691 OcclusionIllumination Viewpoint (a) Large intra -class variation (b) Small intra -class variationRmaxRmax Figure 2. Different levels of intra-class variations in person ReID datasets. (a) Large intra-class variation caused by intensive oc- clusion, illumination changes, and different viewpoints. In this situation, mining the hardest positive pair, which forms the radius of the largest intra-class hypersphere, is harmful to metric learn- ing, while the least-hard pair that shapes the radius of the smallest hypersphere is alternatively an appropriate one. (b) Small intra- class variation with highly similar visual features. The instances in both the hardest and the least-hard pairs share noticeable visual similarities in this case. are either trivial or harmful pairs. To this end, we propose a novel loss paradigm termed Sparse Pairwise (SP) loss that only leverages the most informative pairs for each class in a mini-batch without relying on dense instance-level anchors, as shown in Fig. 1 (bottom row). For the object ReID tasks, we empirically discover that using only few pairs of each class is sufficient for the loss computation as long as the appropriate ones are mined. Based on the proposed loss framework, we conduct a rigorous investigation of the pos- itive mining strategy. Compared to negative mining which has been ac- tively studied [14, 42, 54], positive mining remains under- explored. Generally, an appropriate positive pair should be capable of adapting to different levels of intra-class varia- tions. For example, an occluded pedestrian in a side view- point or dark illumination almost shares no visual similar- ities with his front perspective in Fig. 2a, while the in- stance (blue bounding box) in a clear illumination without any occlusions can bridge them, thus forming an appropri- ate pair for training. To obtain appropriate positive pairs, some works [25, 43, 55] attempt to excavate “moderate” or easy positive pairs, yet still rely on dense anchors. In this work, we first propose a least-hard positive min- ing strategy to address large intra-class variations. Inspired by a geometric insight, we find that the hardest positive pair in a class shapes the radius of the largest hypersphere cov- ering all intra-class samples. Nevertheless, it can be exces- sively large and thus introduces overwhelming visual differ- ences on account of large intra-class variations. In this case, the least-hard positive pair that constructs the smallest hy- persphere can be utilized instead as a more appropriate onesharing noticeable visual similarities (Fig. 2a). In addition, to handle classes with diverse intra-class variations (Fig. 2), we develop an adaptive mining approach that automatically reweights these two kinds of positive pairs, which shapes a dynamic intra-class hypersphere according to the particular situation of the class. The main contributions of this paper are summarized as follows: • We propose a novel pairwise loss framework – Sparse Pairwise loss – that only leverages few informative positive/negative pairs for each class in a mini-batch. • We propose a least-hard positive mining strategy to ad- dress large intra-class variations and further endow it with an adaptive mechanism according to different lev- els of intra-class variations in each mini-batch. • The proposed AdaSP loss is evaluated on various per- son/vehicle ReID datasets and outperforms existing dense pairwise losses across different benchmarks.
Zhao_Representation_Learning_for_Visual_Object_Tracking_by_Masked_Appearance_Transfer_CVPR_2023	Abstract Visual representation plays an important role in visual object tracking. However, few works study the tracking- specified representation learning method. Most trackers directly use ImageNet pre-trained representations. In this paper, we propose masked appearance transfer, a simple but effective representation learning method for tracking, based on an encoder-decoder architecture. First, we en- code the visual appearances of the template and search region jointly, and then we decode them separately. Dur- ing decoding, the original search region image is recon- structed. However, for the template, we make the decoder reconstruct the target appearance within the search region. By this target appearance transfer, the tracking-specified representations are learned. We randomly mask out the inputs, thereby making the learned representations more dis- criminative. For sufficient evaluation, we design a simple and lightweight tracker that can evaluate the representa- tion for both target localization and box regression. Exten- sive experiments show that the proposed method is effec- tive, and the learned representations can enable the simple tracker to obtain state-of-the-art performance on six datasets. https://github.com/difhnp/MAT	1. Introduction Visual object tracking is a computer vision task that highly depends on the quality of visual representation [38]. On this basis, deep representations are adopted for track- ing and successfully boost the development of tracking al- gorithms in previous years. Unlike some primitive track- ers (e.g., [10, 28, 37]) that use ready-made deep features, SiamFC [1] integrates a convolutional neural network (CNN) into the tracking model and learns task-specified representa- tions by end-to-end tuning. From here on, end-to-end model training becomes a common practice in siamese-based track- ing methods. The assumption of the siamese tracker is that *Corresponding author: Huchuan Lu, [email protected] Encoder DecoderImage Image(a) Masked autoencoder. Encoder Decoder DecoderSearch region Search region Template Appearance in SearchShare weights (b) Masked appearance transfer. Figure 1. Comparison between the masked autoencoder [17] and our masked appearance transfer that uses a nontrivial learning objective. different visual appearances of the same target can be em- bedded to have similar representations, and tracking can be achieved by similarity matching. Based on this assumption, the learning objective aims to push the representations of the same target to be close to each other. Although we have a clear learning objective, research on the effective representation learning methods for siamese tracking is still lacking. The common practice is to sim- ply fine-tune the ImageNet [33] pre-trained representations. However, they are learned for classification rather than track- ing. High ImageNet accuracy does not indicate good per- formance in visual object tracking [43]. Without good rep- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 18696 resentations, simple similarity matching cannot bring good tracking performance. Thus, most works focus on the design of tracker neck [7, 36, 47, 50] and tracker head [25, 40, 46]. Within these works, we can notice that the entire model is always trained end-to-end, and representation learning is cou- pled to the box regression task and target localization task. It means that the representation learning is totally driven by other tracker components. To improve tracking performance, we have to make the neck or head increasingly complex. On the aforementioned observation, we attempt to im- prove the tracking performance by learning good representa- tions and propose a simple but effective representation learn- ing method in this paper. This method aims to decouple the representation learning from tracker training, thereby mak- ing it a tracker-free method. It employs a transformer-based autoencoder to learn more discriminative visual features for object tracking. This is achieved by setting a nontrivial learning objective for the autoencoder, as illustrated by Fig- ure 1b. The embedded target appearance of the template can be pushed to be close to that in the search region. We further make the learned representations more discriminative by masking out the input images. To evaluate the learned representations for target localiza- tion and box regression, we design a very simple tracker with a matching operator and a lightweight head. We evaluate the proposed method with different model architectures and initial weights. We also compare our simple tracker with many state-of-the-art trackers on the popular LaSOT [14], TrackingNet [31], GOT10k [20], and many other tracking datasets [15, 30, 41]. The experiments demonstrate the effec- tiveness and generalization ability of our proposed represen- tation learning method. Comparison results show that our simple tracker can obtain state-of-the-art performance with the learned representations. To summarize, our contributions are presented as follows: •We propose masked appearance transfer (MAT), a novel representation learning method for visual object track- ing that jointly encodes the template and search region images, and learns tracking-specified representation with a simple encoder-decoder pipeline. A nontrivial training objective is proposed to make this method ef- fective. •We design a simple and lightweight tracker for tracking- specified evaluation that has few parameters and no hyper-parameters. It can evaluate the representations not only for target localization but also for box regres- sion. •Extensive experiments demonstrate the effectiveness and generalization ability of the proposed method. Ex- tensive comparisons show that the proposed simple tracker can obtain state-of-the-art performance by us- ing the learned tracking-specified representations.
Zhou_Unsupervised_Cumulative_Domain_Adaptation_for_Foggy_Scene_Optical_Flow_CVPR_2023	Abstract Optical ﬂow has achieved great success under clean scenes, but suffers from restricted performance under foggy scenes. To bridge the clean-to-foggy domain gap, the ex- isting methods typically adopt the domain adaptation to transfer the motion knowledge from clean to synthetic foggy domain. However, these methods unexpectedly neglect the synthetic-to-real domain gap, and thus are erroneous when applied to real-world scenes. To handle the practical optical ﬂow under real foggy scenes, in this work, we propose a novel unsupervised cumulative domain adaptation optical ﬂow (UCDA-Flow) framework: depth-association motion adaptation and correlation-alignment motion adaptation. Speciﬁcally, we discover that depth is a key ingredient to in- ﬂuence the optical ﬂow: the deeper depth, the inferior optical ﬂow, which motivates us to design a depth-association mo- tion adaptation module to bridge the clean-to-foggy domain gap. Moreover, we ﬁgure out that the cost volume correlation shares similar distribution of the synthetic and real foggy im- ages, which enlightens us to devise a correlation-alignment motion adaptation module to distill motion knowledge of the synthetic foggy domain to the real foggy domain. Note that synthetic fog is designed as the intermediate domain. Under this uniﬁed framework, the proposed cumulative adaptation progressively transfers knowledge from clean scenes to real foggy scenes. Extensive experiments have been performed to verify the superiority of the proposed method.	1. Introduction Optical ﬂow has made great progress under clean scenes, but may suffer from restricted performance under foggy scenes [15]. The main reason is that fog weakens scene contrast, breaking the brightness and gradient constancy assumptions, which most optical ﬂow methods rely on. To alleviate this, researchers start from the perspective *Corresponding author. fog gap style gap Clean Scene Source ğ Syn. Fog Scene Intermediate ğ Real Fog Scene Target ğ Target Domain Source Domain Correlation Alignment AdaptationIntermediate DomainDepth Association Adaptation Association Depth / Image / Flow / Correlation Figure 1. Illustration of the main idea. We propose to transfer motion knowledge from the source domain (clean scene) to the target domain (real foggy scene) through two-stage adaptation. We design the synthetic foggy scene as the intermediate domain. As for the clean-to-foggy domain gap (fog), we transfer motion knowledge from the source domain to the intermediate domain via depth association. As for the synthetic-to-real domain gap (style), we distill motion knowledge of the intermediate domain to the target domain by aligning the correlation of both the domains. of domain adaptation, which mainly seeks the degradation- invariant features to transfer the motion knowledge from the clean scene to the adverse weather scene [14 –16, 40]. For example, Li [15,16] attempted to learn degradation-invariant features to enhance optical ﬂow under rainy scenes in a su- pervised manner. Yan et al. [40] proposed a semi-supervised framework for optical ﬂow under dense foggy scenes, which relies on the motion-invariant assumption between the paired clean and synthetic foggy images. These pioneer works have made a good attempt to handle the clean-to-foggy domain gap with synthetic degraded images through one-stage do- main adaptation. However, they lack the constraints to guide the network to learn the motion pattern of real foggy domain, and fail for real foggy scenes. In other words, they have un- expectedly neglected the synthetic-to-real domain gap, thus limiting their performances on real-world foggy scenes. In this work, our goal is to progressively handle the two domain gaps: the clean-to-foggy gap and the synthetic-to-real gap in a cumulative domain adaptation framework in Fig. 1. This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 9569 As for the clean-to-foggy gap, we discover that depth is a key ingredient to inﬂuence the optical ﬂow: the deeper the depth, the inferior the optical ﬂow. This observation inspires us to explore the usage of depth as the key to bridging the clean-to-foggy gap (seeing the fog gap in Fig. 1). On one hand, depth physically associates the clean image with the foggy image through atmospheric scattering model [26]; on the other hand, there exists a natural 2D-3D geometry pro- jection relationship between depth and optical ﬂow, which is used as a constraint to transfer motion knowledge from the clean domain to the synthetic foggy domain. As for the synthetic-to-real gap, we ﬁgure out that cost volume correlation shares similar distribution of synthetic and real foggy images. Cost volume stores correlation value, which can physically measure the similarity between adja- cent frames, regardless of synthetic and real foggy images. Therefore, cost volume beneﬁts to bridging the synthetic-to- real domain gap (seeing the style gap in Fig. 1). We align the correlation distributions to distill motion knowledge of the synthetic foggy domain to the real foggy domain. In this work, we propose a novel unsupervised cumulative domain adaptation optical ﬂow (UCDA-Flow) framework for real foggy scene, including depth-association motion adapta- tion (DAMA) and correlation-alignment motion adaptation (CAMA). Speciﬁcally, in DAMA stage, we ﬁrst estimate optical ﬂow, ego-motion and depth with clean stereo images, and then project depth into optical ﬂow space with 2D-3D geometry formula between ego-motion and scene-motion to enhance rigid motion. To bridge the clean-to-foggy gap, we utilize atmospheric scattering model [26] to synthesize the corresponding foggy images, and then transfer motion knowledge from the clean domain to the synthetic foggy do- main. In CAMA stage, to bridge the synthetic-to-real domain gap, we transform the synthetic and real foggy images to the cost volume space, in which we align the correlation distri- bution to distill the motion knowledge of the synthetic foggy domain to the real foggy domain. The proposed cumulative domain adaptation framework could progressively transfer motion knowledge from clean domain to real foggy domain via depth association and correlation alignment. Overall, our main contributions are summarized as follows: •We propose an unsupervised cumulative domain adapta- tion framework for optical ﬂow under real foggy scene, consisting of depth-association motion adaptation and correlation-alignment motion adaptation. The proposed method can transfer motion knowledge from clean domain to real foggy domain through two-stage adaptation. •We reveal that foggy scene optical ﬂow deteriorates with depth. The geometry relationship between depth and opti- cal ﬂow motivates us to design a depth-association motion adaptation to bridge the clean-to-foggy domain gap. •We illustrate that cost volume correlation distribution of the synthetic and real foggy images is consistent. Thisprior beneﬁts to close the synthetic-to-real domain gap through correlation-alignment motion adaptation.
Zhou_NeRFLix_High-Quality_Neural_View_Synthesis_by_Learning_a_Degradation-Driven_Inter-Viewpoint_CVPR_2023	Abstract Neural radiance fields (NeRF) show great success in novel view synthesis. However, in real-world scenes, re- covering high-quality details from the source images is still challenging for the existing NeRF-based approaches, due to the potential imperfect calibration information and scene representation inaccuracy. Even with high-quality train- ing frames, the synthetic novel views produced by NeRF models still suffer from notable rendering artifacts, such as noise, blur, etc. Towards to improve the synthesis qual- ity of NeRF-based approaches, we propose NeRFLiX, a general NeRF-agnostic restorer paradigm by learning a degradation-driven inter-viewpoint mixer. Specially, we de- *Equal contribution †Corresponding authorsign a NeRF-style degradation modeling approach and con- struct large-scale training data, enabling the possibility of effectively removing NeRF-native rendering artifacts for ex- isting deep neural networks. Moreover, beyond the degra- dation removal, we propose an inter-viewpoint aggregation framework that is able to fuse highly related high-quality training images, pushing the performance of cutting-edge NeRF models to entirely new levels and producing highly photo-realistic synthetic views.	1. Introduction Neural radiance fields (NeRF) can generate photo- realistic images from new viewpoints, playing a heated role in novel view synthesis. In light of NeRF’s [37] success, numerous approaches [2, 9, 11, 19, 34, 35, 38, 40, 41, 47, 53, 54,59,61,65] along these lines have been proposed, contin- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 12363 ually raising the performance to greater levels. In fact, one prerequisite of NeRF is the precise camera settings of the taken photos for training [22, 32, 61]. However, accurately calibrating camera poses is exceedingly difficult in prac- tice. Contrarily, the shape-radiance co-adaption issue [74] reveals that while the learned radiance fields can perfectly explain the training views with inaccurate geometry, they poorly generalize to unseen views. On the other hand, the capacity to represent sophisticated geometry, lighting, ob- ject materials, and other factors is constrained by the simpli- fied scene representation of NeRF [19,77,78]. On the basis of such restrictions, advanced NeRF models may nonethe- less result in notable artifacts (such as blur, noise, detail missing, and more), which we refer to as NeRF-style degra- dations in this article and are shown in Fig. 1. To address the aforementioned limitations, numerous works have been proposed. For example, some studies, including [22, 59, 66, 70], jointly optimize camera param- eters and neural radiance fields to refine camera poses as precisely as possible in order to address the camera calibra- tion issue. Another line of works [19, 73, 77, 78] presents physical-aware models that simultaneously take into ac- count the object materials and environment lighting, as op- posed to using MLPs or neural voxels to implicitly encode both the geometry and appearance. To meet the demands for high-quality neural view synthesis, one has to carefully examine all of the elements when building complex inverse rendering systems. In addition to being challenging to op- timize, they are also not scalable for rapid deployment with hard re-configurations in new environments. Regardless of the intricate physical-aware rendering models, is it possi- ble to design a practical NeRF-agnostic restorer to directly enhance synthesized views from NeRFs ? In the low-level vision, it is critical to construct large- scale paired data to train a deep restorer for eliminating real-world artifacts [56, 72]. When it comes to NeRF-style degradations, there are two challenges: (1) sizable paired training data; (2) NeRF degradation analysis. First, it is un- practical to gather large-scale training pairs (more specifi- cally, raw outputs from well-trained NeRFs and correspond- ing ground truths). Second, the modeling of NeRF-style degradation has received little attention. Unlike real-world images that generally suffer from JPEG compression, sen- sor noise, and motion blur, the NeRF-style artifacts are complex and differ from the existing ones. As far as we know, noprevious studies have ever investigated NeRF- style degradation removal which effectively leverages the ample research on image and video restoration. In this work, we are motivated to have the firststudy on the feasibility of simulating large-scale NeRF-style paired data, opening the possibility of training a NeRF-agnostic restorer for improving the NeRF rendering frames. To this end, we present a novel degradation simulator for typ-ical NeRF-style artifacts (e.g., rendering noise and blur) considering the NeRF mechanism. We review the over- all NeRF rendering pipeline and discuss the typical NeRF- style degradation cases. Accordingly, we present three basic degradation types to simulate the real rendered artifacts of NeRF synthetic views and empirically evaluate the distri- bution similarity between real rendered photos and our sim- ulated ones. The feasibility of developing NeRF-agnostic restoration models has been made possible by constructing a sizable dataset that covers a variety of NeRF-style degra- dations, over different scenes. Next, we show the necessity of our simulated dataset and demonstrate that existing state-of-the-art image restoration frameworks can be used to eliminate NeRF visual artifacts. Furthermore, we notice, in a typical NeRF setup, neigh- boring high-quality views come for free, and they serve as potential reference bases for video-based restoration with a multi-frame aggregation and fusion module. However, this is not straightforward because NeRF input views are taken from a variety of very different angles and locations, mak- ing the estimation of correspondence quite challenging. To tackle this problem, we propose a degradation-driven inter- viewpoint “mixer” that progressively aligns image contents at the pixel and patch levels. In order to maximize efficiency and improve performance, we also propose a fast view se- lection technique to only choose the most pertinent refer- ence training views for aggregation, as opposed to using the entire NeRF input views. In a nutshell, we present a NeRF-agnostic restorer (termed NeRFLiX) which learns a degradation-driven inter- viewpoint mixer. As illustrated in Fig. 1, given NeRF syn- thetic frames with various rendering degradations, NeR- FLiX successfully restores high-quality results. Our con- tributions are summarized as •Universal enhancer for NeRF models. NeRFLiX is powerful and adaptable, removing NeRF artifacts and restoring clearly details, pushing the performance of cutting-edge NeRF models to entirely new levels. •NeRF rendering degradation simulator. We develop a NeRF-style degradation simulator (NDS), construct- ing massive amounts of paired data and aiding the training of deep neural networks to improve the quality of NeRF-rendered images. •Inter-viewpoint mixer. Based on our constructed NDS, we further propose an inter-viewpoint baseline that is able to mixhigh-quality neighboring views for more effective restorations. •Training time acceleration. We show how NeRFLiX makes it possible for NeRF models to produce even better results with a 50%reduction in training time. 12364
Zhu_IPCC-TP_Utilizing_Incremental_Pearson_Correlation_Coefficient_for_Joint_Multi-Agent_Trajectory_CVPR_2023	Abstract Reliable multi-agent trajectory prediction is crucial for the safe planning and control of autonomous systems. Com- pared with single-agent cases, the major challenge in si- multaneously processing multiple agents lies in modeling complex social interactions caused by various driving in- tentions and road conditions. Previous methods typically leverage graph-based message propagation or attention mechanism to encapsulate such interactions in the format of marginal probabilistic distributions. However, it is in- herently sub-optimal. In this paper, we propose IPCC- TP , a novel relevance-aware module based on Incremental Pearson Correlation Coefficient to improve multi-agent in- teraction modeling. IPCC-TP learns pairwise joint Gaus- sian Distributions through the tightly-coupled estimation of the means and covariances according to interactive incre- mental movements. Our module can be conveniently em- bedded into existing multi-agent prediction methods to ex- tend original motion distribution decoders. Extensive ex- periments on nuScenes and Argoverse 2 datasets demon- strate that IPCC-TP improves the performance of baselines by a large margin.	1. Introduction Trajectory prediction refers to predicting the future tra- jectories of one or several target agents based on past trajec- tories and road conditions. It is an essential and emerging subtask in autonomous driving [22,27,36,44] and industrial robotics [20, 34, 47]. Previous methods [6, 14, 17, 32, 48] concentrate on Single-agent Trajectory Prediction (STP), which leverages past trajectories of other agents and surrounding road condi- tions as additional cues to assist ego-motion estimation. De- ∗Equal contribution.†Corresponding author. ①②③④⑤-101!①②!②③!③④!④⑤ ①②③④⑤(a)(b) !!"!#"!MeansCovariances!#$!Periodoftime"!#-thdistribution"#$Probabilitydensity$%Displacement$%Figure 1. Schematic illustration of a driving scene with five agents ①−⑤.(a)IPCC-TP models future interactions for tsteps based on joint Gaussian distribution among one-dimensional increments of agents’ displacement d.(b)The values of IPCC represented by ρ on the axis indicate the correlation between agents’ trajectories. In the predicted future, ①and②interact intensively with ①yielding to②(ρ12<0).④’s movement does not interfere ⑤(\|ρ45\| ≈0). ③is closely following ④(ρ34≤+1). spite the significant progress made in recent years, its appli- cation is limited. The reason is that simultaneously process- ing all traffic participants remains unreachable, although it is a common requirement for safe autonomous driving. A straightforward idea to deal with this issue is to di- rectly utilize STP methods on each agent respectively and finally attain a joint collision-free prediction via pairwise collision detection. Such methods are far from reliably handling Multi-agent Trajectory Prediction (MTP) since the search space for collision-free trajectories grows exponen- tially as the number of agents increases, i.e.,Nagents with Mmodes yield MNpossible combinations, making such search strategy infeasible with numerous agents. By in- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 5507 corporating graph-based message propagation or attention mechanism into the framework to model the future interac- tions among multiple agents, [16,30,42] aggregate all avail- able cues to simultaneously predict the trajectories for mul- tiple agents in the format of a set of marginal probability distributions. While these frameworks improve the origi- nal STP baseline to some extent, the estimation based on a set of marginal distributions is inherently sub-optimal [33]. To address this issue, a predictor that can predict a joint probability distribution for all agents in the scene is neces- sary. An intuitive strategy for designing such a predictor is to directly predict all parameters for means and covari- ance matrices, which define the joint probability distribu- tions [38]. To predict a trajectory of Ttime-steps with N agents and Mmodes up to 4N2TM covariance parame- ters are needed. However, such redundant parameterization fails to analyze the physical meaning of covariance parame- ters and also makes it hard to control the invertibility of the covariance matrix. In this paper, we take a further step to explicitly inter- pret the physical meanings of the individual variances and Pearson Correlation Coefficients (PCC) inside the covari- ance matrix. Instead of directly modeling the locations of agents at each sampled time, we design a novel movement- based method which we term Incremental PCC (IPCC). It helps to simulate the interactions among agents during driv- ing. IPCC models the increment from the current time to a future state at a specific step. Thereby individual variances capture the local uncertainty of each agent in the Bird-Eye- View without considering social interactions (Figure 1.a), while IPCC indicates the implicit driving policies of two agents, e.g., one agent follows or yields to another agent or the two agents’ motions are irrelevant (Figure 1.b). Com- pared to position-based modeling, IPCC has two advan- tages. First, IPCC models social interactions during driving in a compact manner. Second, since position-based model- ing requires respective processes along the two axes, IPCC can directly model the motion vector, which reduces mem- ory cost for covariance matrix storage by a factor of four and facilitates deployment on mobile platforms, such as au- tonomous vehicles and service robots. Based on IPCC, we design and implement IPCC-TP, a module for the MTP task. IPCC-TP can be conveniently embedded into existing methods by extending the original motion decoders to boost their performance. Experiments show that IPCC-TP effectively captures the pairwise rele- vance of agents and predicts scene-compliant trajectories. In summary, the main contributions of this work are threefold: • We introduce a novel movement-based method, namely IPCC, that can intuitively reveal the physical meaning of pairwise motion relevance and facilitate deployment by reducing memory cost.• Based on IPCC, we propose IPCC-TP, which is com- patible with state-of-the-art MTP methods and possess the ability to model future interactions with joint Gaus- sian distributions among multiple agents. • Experiment results show that methods enhanced with IPCC-TP outperform the original methods by a large margin and become the new state-of-the-art methods.
Zhu_Transductive_Few-Shot_Learning_With_Prototype-Based_Label_Propagation_by_Iterative_Graph_CVPR_2023	Abstract Few-shot learning (FSL) is popular due to its ability to adapt to novel classes. Compared with inductive few-shot learning, transductive models typically perform better as they leverage all samples of the query set. The two exist- ing classes of methods, prototype-based and graph-based, have the disadvantages of inaccurate prototype estimation and sub-optimal graph construction with kernel functions, respectively. In this paper, we propose a novel prototype- based label propagation to solve these issues. Specifically, our graph construction is based on the relation between prototypes and samples rather than between samples. As prototypes are being updated, the graph changes. We also estimate the label of each prototype instead of considering a prototype be the class centre. On mini-ImageNet, tiered- ImageNet, CIFAR-FS and CUB datasets, we show the pro- posed method outperforms other state-of-the-art methods in transductive FSL and semi-supervised FSL when some un- labeled data accompanies the novel few-shot task.	1. Introduction With the availability of large-scale datasets and the rapid development of deep convolutional architectures, super- vised learning exceeds in computer vision, voice, and ma- chine translation [23]. However, lack of data makes the ex- isting supervised models fail during the inference on novel tasks. As the annotation process may necessitate expert knowledge, annotations are may be scarce and costly ( e.g., annotation of medical images). In contrast, humans can learn a novel concept from just a single example. Few-shot learning (FSL) aims to mimic the capabilities of biological vision [7] and it leverages metric learning, meta-learning, or transfer learning. The purpose of metric- based FSL is to learn a mapping from images to an embed- ding space in which images from the same class are closer *The corresponding author. Code: https : / / github . com / allenhaozhu/protoLP Figure 1. Drawbacks of prototype-based and graph-based FSL. (left) Some label assignments are incorrect due to the imperfect decision boundary. ( right ) Some “strong” links in the fixed graph are incorrect as they associate samples of di fferent classes. together and images from other classes are separated. Meta- learning FSL performs task-specific optimisation with the goal to generalize to other tasks well. Pre-training a fea- ture extractor followed by adapting it for reuse on new class samples is an example of transfer learning. Several recent studies [6, 11, 13, 16, 22, 29, 34, 35] ex- plored transductive inference for few-shot learning. At the test time, transductive FSL infer the class label jointly for all the unlabeled query samples, rather than for one sam- ple/episode at a time. Thus, transductive FSL typically out- performs inductive FSL. We categorise transductive FSL into: (i) FSL that requires the use of unlabeled data to esti- mate prototypes [2,26,27,43,44], and (ii) FSL that builds a graph with some kernel function and then uses label prop- agation to predict labels on query sets [22, 29, 61]. How- ever, the above two paradigms have their own drawbacks. For prototype-based methods, they usually use the nearest neighbour classifier, which is based on the assumption that there exists an embedding where points cluster around a single prototype representation for each class. Fig. 1 (left) shows a toy example which is sensitive to the large within- class variance and low between-class variance. Thus, the This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 23996 two prototypes cannot be estimated perfectly by the soft label assignment alone. Fig. 1 (right) shows that Label- Propagation (LP) and Graph Neural Network (GNN) based methods depend on the graph construction which is com- monly based on a specific kernel function determining the final result. If some nodes are wrongly and permanently linked, these connections will a ffect the propagation step. In order to avoid the above pitfalls of transductive FSL, we propose prototype-based Label-Propagation (protoLP). Our transductive inference can work with a generic feature embedding learned on the base classes. Fig. 2 shows how to alternatively optimize a partial assignment between pro- totypes and the query set by (i) solving a kernel regression problem (or optimal transport problem) and (ii) a label prob- ability prediction by prototype-based label propagation. Im- portantly, protoLP does not assume the uniform class distri- bution prior while significantly outperforming other meth- ods that assume the uniform prior, as shown in ablations on the imbalanced benchmark [46] where methods relying on the balanced class prior fail. Our model outperforms state- of-the-art methods significantly, consistently providing im- provements across di fferent settings, datasets, and training models. Our transductive inference is very fast, with run- times that are close to the runtimes of inductive inference. Our contributions are as follows: i. We identify issues resulting from separation of proto- type-based and label propagation methods. We propose prototype-based Label Propagation (protoLP) for trans- ductive FSL, which unifies both models into one frame- work. Our protoLP estimates prototypes not only from the partial assignment but also from the prediction of la- bel propagation. The graph for label propagation is not fixed as we alternately learn prototypes and the graph. ii. By introducing parameterized label propagation step, we remove the assumption of uniform class prior while other methods highly depend on this prior. iii. We showcase advantages of protoLP on four datasets for transductive and semi-supervised learning, Our pro- toLP outperforms the state of the art under various set- tings including di fferent backbones, unbalanced query set, and data augmentation.
Zhou_Relightable_Neural_Human_Assets_From_Multi-View_Gradient_Illuminations_CVPR_2023	Abstract Human modeling and relighting are two fundamental problems in computer vision and graphics, where high- quality datasets can largely facilitate related research. However, most existing human datasets only provide multi- view human images captured under the same illumination. Although valuable for modeling tasks, they are not read- ily used in relighting problems. To promote research in both fields, in this paper, we present UltraStage, a new 3D human dataset that contains more than 2,000 high-quality human assets captured under both multi-view and multi- illumination settings. Specifically, for each example, we provide 32 surrounding views illuminated with one white light and two gradient illuminations. In addition to regular multi-view images, gradient illuminations help recover de- tailed surface normal and spatially-varying material maps, enabling various relighting applications. Inspired by re- *Equal contribution. †Corresponding author.cent advances in neural representation, we further interpret each example into a neural human asset which allows novel view synthesis under arbitrary lighting conditions. We show our neural human assets can achieve extremely high capture performance and are capable of representing fine details such as facial wrinkles and cloth folds. We also validate UltraStage in single image relighting tasks, training neural networks with virtual relighted data from neural assets and demonstrating realistic rendering improvements over prior arts. UltraStage will be publicly available to the community to stimulate significant future developments in various hu- man modeling and rendering tasks. The dataset is available at https://miaoing.github.io/RNHA.	1. Introduction Multi-view stereo (MVS) and photometric stereo (PS) have long served as two complementary workhorses for re- covering 3D objects, human performances, and environ- ments [1,29]. Earlier MVS typically exploits feature match- ing and bundle adjustment to find ray correspondences This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 4315 across varying viewpoints [13, 43, 82] and subsequently infer their corresponding 3D points [19, 72, 81]. More re- cent neural modeling approaches have emerged as a more effective solution by implicitly encoding both geometry and appearance using neural networks [58]. PS, in contrast, generally assumes a single (fixed) viewpoint and employs appearance variations under illumination changes to infer the surface normal and reflectance (e.g., albedo) [1,20,23– 25, 35, 53, 54, 62]. Shape recovery in PS essentially corre- sponds to solving inverse rendering problems [25] where recent approaches also move towards neural representa- tions to encode the photometric information [55,56,68,90]. Most recently, in both MVS and PS neural representations have demonstrated reduced data requirement [48, 88] and increased accuracy [21, 22]. In the context of 3D human scanning, MVS and PS exhibit drastically different benefits and challenges, in synchronization, calibration, reconstruc- tion, etc. For example, for high-quality performance cap- ture, MVS has long relied on synchronized camera arrays but assumes fixed illumination. The most popular and per- haps effective apparatus is the camera dome with tens and even hundreds of cameras [50]. Using classic or neural rep- resentations, such systems can produce geometry with rea- sonable quality. However, ultra-fine details such as facial wrinkles or clothing folds are generally missing due to lim- ited camera resolutions, small camera baselines, calibration errors, etc [87]. In comparison, a typical PS capture system uses a sin- gle camera and hence eliminates the need for cross-camera synchronization and calibrations. Yet the challenges shift to synchronization across the light sources and between the lights and the camera, as well as calibrating the light sources. PS solutions are epitomized by the USC Light- Stage [15, 30, 87] that utilize thousands of light sources to provide controllable illuminations [23,25,53,54,79], with a number of recent extensions [2,27,34,35,41,65,73,95,96]. A key benefit of PS is that it can produce very high-quality normal maps significantly surpassing MVS reconstruction. Further, the appearance variations can be used to infer sur- face materials and conduct high-quality appearance editing and relighting applications [2, 65, 79, 95, 96]. However, re- sults from a single camera cannot fully cover the complete human geometry. Nor have they exploited multi-view re- constructions as useful priors. A unified capture apparatus that combines MVS and PS reconstructions has the poten- tial to achieve unprecedented reconstruction quality, rang- ing from recovering ultra-fine details such as clothes folds and facial wrinkles to supporting free-view relighting in metaverse applications. In particular, the availability of a comprehensive MVS-PS dataset may enable new learning- based approaches for reconstruction [58,66,77,89,92,100], rendering [11,58,83,86,93,98], and generation [18,37–39, 71]. However, there is very little work or dataset available Figure 2. Data overview. Here we show 5 examples in our dataset. From front to back are the normal map, albedo map, color gradient illumination observation, and inverse color gradient illumination observation, respectively. In total, UltraStage Dataset contains more than 2,000 human-centric scenes of a single person or multi- ple people with various gestures, clothes and interactions. For each scene, we provide 8K resolution images captured by 32 surround- ing cameras under 3 illuminations. See Supp. for more examples. to the public due to challenges on multiple fronts for con- structing such imaging systems. To fill in the gap, we construct PlenOptic Stage Ultra, an emerging hardware system that conducts simultaneous MVS and PS acquisition of human performances. PlenOp- tic Stage Ultra is built upon an 8-meter radius large-scale capture stage with 22,080 light sources to illuminate a per- former with controllable illuminations, as well as places 32 cameras on the cage to cover the 360◦surrounding views of the object. We present detailed solutions to obtain accurate camera-camera and camera-light calibration and synchro- nization as well as conduct camera ISP correction and light- ing intensity rectification. For PS capture, PlenOptic Stage Ultra adopts tailored illuminations [27,53,55,57]: for each human model, we illuminate it with one white light and two directional gradient illuminations. This produces ex- tra high-quality surface normal and anisotropic reflectance largely missing in existing MVS human reconstruction. A direct result of our system is UltraStage, a novel human dataset that provides more than 2,000 human models under different body movements and with sophisticated clothing like frock or cheongsam, to be disseminated to the commu- nity. We further demonstrate several neural modeling and ren- dering techniques [11, 58, 59, 93, 98, 99] to process the dataset for recovering ultra-fine geometric details, modeling surface reflectance, and supporting relighting applications. Specifically, we show neural human assets achieve signifi- cantly improved rendering and reconstruction quality over purely MVS or PS based methods [6,19,20,25,53,72], e.g., they can render exquisite details such as facial wrinkles and cloth folds. To use the neural assets for relighting, we adopt albedo and normal estimation networks for in-the-wild hu- man full-body images and we show our dataset greatly en- hances relighting quality over prior art [36,80]. The dataset as well as the processing tools are expected to stimulate significant future developments such as generation tasks in both shape and appearance. 4316 Figure 3. System Overview. The PlenOptic Stage Ultra is an 8- meter lighting stage composed of 32 sounding cameras and 22,080 controllable light sources. It supports both MVS and PS capture settings, enabling high-quality geometry and material acquisition for large-scale subjects and objects. Note that a human is inside in the middle of the picture.
Zheng_Curricular_Contrastive_Regularization_for_Physics-Aware_Single_Image_Dehazing_CVPR_2023	Abstract Considering the ill-posed nature, contrastive regular- ization has been developed for single image dehazing,introducing the information from negative images as alower bound. However , the contrastive samples are non-consensual, as the negatives are usually represented dis-tantly from the clear (i.e., positive) im age, leaving the so- lution space still under-constricted. Moreover , the inter-pretability of deep dehazing models is underexplored to-wards the physics of the hazing process. In this paper , wepropose a novel curricular contrastive regularization tar-geted at a consensual contrastive space as opposed to anon-consensual one. Our negatives, which provide betterlower-bound constraints, can be assembled from 1) the hazyimage, and 2) corresponding restorations by other existing methods. Further , due to the different similarities betweenthe embeddings of the clear image and negatives, the learn-ing difﬁculty of the multiple components is intrinsically im-balanced. To tackle this issue, we customize a curriculum learning strategy to reweight the importance of differentnegatives. In addition, to improve the interpretability in the feature space, we build a physics-aware dual-branchunit according to the atmospheric scattering model. Withthe unit, as well as curricular contrastive regularization, we establish our dehazing network, named C 2PNet. Ex- tensive experiments demonstrate that our C2PNet signiﬁ- cantly outperforms state-of-the-art methods, with extremePSNR boosts of 3.94dB and 1.50dB, respectively, on SOTS-indoor and SOTS-outdoor datasets. Code is available athttps://github.com/YuZheng9/C2PNet .	1. Introduction As a common atmospheric phenomenon, haze noticeably degrades the quality of photographed images, severely lim- iting the performance of subsequent high-level visual taskssuch as vehicle re-identiﬁcation [ 7] and scene understand- *Corresponding author ([email protected]). Positive(GT) Ultra-hard Negatives Easy Negatives (g) Rate E:H:U = 1:1:1 Hard Negatives NegativesAnchor Non-consensual Contrastive Space Consensual Contrastive Space × Push Pull × Push (under-constricted)#Negatives in the consensual contrastive spacePSNR (dB) 3335373839 3436 01 3 (a) 36.39(b) 37.89 (c) 33.96(e) 38.57(f) 38.25 (g) 38.68 (d) 38.19 Method (a) Baseline (b) (c) (d) (e) (f) (g)Rate E:H:U 0:0:0 1:0:0 0:0:10:1:0 1:2:0 1:0:21:1:1 Figure 1. Upper panel: Examination for contrastive regularization based on three difﬁculty levels of the negatives in the consensualcontrastive space. Lower panel: Illustration of contrastive samplesin the consensual and non-consensual spaces. ing [ 35]. Similar to the emergence of other image restora- tion task solvers [ 12,13,39,43], valid image dehazing tech- niques are required for handling vision-based applications. Deep learning based methods have achieved tremendous success in single image dehazing and can be roughly cate- gorized into two classes: physics-free methods [ 5,10,17,24] and physics-aware methods [ 4,8,11,34]. Regarding the for- mer, most of them usually use ground-truth images withpredicted restorations to enforce L1/L2 distance-based con-sistency and also involve various regularizations [ 29,42]a s additional constraints to cope with the ill-posed property. Notice that all of those regularizations ignore the informa-tion from negative images as a lower bound, contrastive reg-ularization (CR) [ 40] is proposed to introduce different hazy This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 5785 images as negatives and the ground-truth image as the pos- itive and further uses contrastive learning [ 19,20] to guar- antee a closed solution space. Moreover, it is shown thatbetter performances can be achieved when using more neg-atives since diverse degraded patterns are included as cues. However, the issue is that the contents of those negatives aredistinct from the positive, and their embeddings may be too distant, leaving the solution space still under-constricted. To remedy this issue, a natural idea is to use the negatives in the consensual contrastive space 1(see the lower panel in Fig. 1) as better lower-bound constraints, which can be easily assembled from the hazy input and the correspondingrestorations by other existing methods. In such cases, thenegatives can be “closer” to the positive than those in thenon-consensual space since the diversity of such negativesis more associated with the haze (or haze residue) ratherthan any other semantics. However, an intrinsic dilemmaarises when the embedding of a negative is too close to thatof the positive, as its pushing force to an anchor ( i.e., the prediction) may cancel out the pulling force of the positive.Such a learning difﬁculty can confuse the anchor to movetowards the positive, especially in the early training stage. This intuition is further examined in the upper panel of Fig. 1. We use FFA-Net [ 33] as baseline (row (a)) and SOTS-indoor [ 28] as the testing dataset to explore the im- pact of the negatives in the consensual space with diversedifﬁculty. Speciﬁcally, we deﬁne the difﬁculty of the neg-atives into three levels: easy (E), hard (H), and ultra-hard (U). We adopt the hazy input as the easy negative, and use a coarse strategy to distinguish between the latter two types,i.e., whether the PSNR of the negative is greater than 30. First, in the single-negative case (row (b)-(d)), an interestingﬁnding is that using a hard sample as negative achieves thebest performance compared to the other two settings, andusing an ultra-hard negative is even worse than the base-line. This reveals that a “close” negative has the potentialto promote the effectiveness of the dehazing model, but notthe closer the better due to the learning difﬁculty. Whilein the multi-negative case 2(row (e)-(g)), we have observed that comprehensively covering negatives with different dif-ﬁculty levels, including ultra-hard samples, can lead to thebest performance. It implies the negatives at different difﬁ-culty levels can all contribute to the training phase. Theseobservations motivate us to explore how to wisely arrangethe multiple negative pairs in a consensual space into theCR during training. Moving on to the realm of physics-aware deep models, 1In this space, the contents of the negatives are identical to the positive sample, except for the haze distribution. Here, we use the terms (non- )consensual contrastive space and (non-)consensual space interchangeably, and a negative in the consensual space is denoted as a consensual negative. 2We give each negative the same weight in the regularization under this case, and we omit the cases of E=0, which would drastically decrease the performance. We will discuss the reason for this in Sec. 3.most of them utilize the atmospheric scattering model [ 31, 32] in the raw space, without fully exploring the beneﬁ- cial feature-level information. PFDN [ 11] is the only work that attempts to express the physics model as a basic unitin the network. The unit is designed as a shared structureto predict the latent features corresponding to the atmo-spheric light and transmission map. Nevertheless, the for-mer is usually assumed to be homogeneous while the latteris non-homogeneous, and thus their features cannot be ap-proximated in the same way. Therefore, it is still an openproblem how to accurately realize the interpretability of thefeature space of the deep network using the physics model,which is another aspect we are interested in. In this paper, we propose a curricular contrastive reg- ularization using hazy or restored images as negatives in the consensual space for image dehazing to address the ﬁrst issue. Informed by our analysis, which suggests that thedifﬁculty of consensual negatives can impact the effective-ness of the regularization, we present a curriculum learningstrategy to arrange these negatives to mitigate learning am-biguity. Speciﬁcally, we split the negatives into three types (i.e., easy, hard, and ultra-hard) and assign different weights to corresponding negative pairs in CR. Meanwhile, the dif-ﬁculty levels of the negatives are dynamically adjusted asthe anchor moves towards the positive in the representationspace during training. In this way, the proposed regular-ization can facilitate the dehazing models to be stably opti- mized in a more compact solution space. We propose a physics-aware dual-branch unit (PDU) re- garding the second issue. The PDU approximates the fea-tures corresponding to the atmospheric light and the trans-mission map in dual branches, respectively considering thephysical characteristics of each factor. The features of thelatent clear image can thus be synthesized more precisely inline with the physics model. Finally, we establish C 2PNet, our dehazing network that deploys PDUs into a cascadedbackbone with curricular contrastive regularization. In summary, our key contributions are as follows: • We propose a novel C 2PNet for haze removal that em- ploys curricular contrastive regularization and enforcesphysics-based prior in the feature space. Our methodoutperforms SOTAs in both synthetic and real-worldscenarios. In particular, we achieve signiﬁcant PSNRboosts of 3.94dB and 1.50dB on the SOTS-indoor andSOTS-outdoor datasets, respectively. • The proposed regularization adopts a unique consen- sual negative-based approach for dehazing and incor-porates a self-contained curriculum learning strategy that dynamically calibrates the priority and difﬁculty levels of the negatives. It is also proven to enhance theperformance of SOTAs as a generalized regularizationtechnique, surpassing previous related strategies. • With careful consideration of the characteristics of fac- 5786 tors involved, we built the PDU based on an unprece- dented expression of the physics model. This innova-tive design promotes feature transmission and extrac-tion in the feature space, guided by physics priors.
Zhu_STMT_A_Spatial-Temporal_Mesh_Transformer_for_MoCap-Based_Action_Recognition_CVPR_2023	Abstract We study the problem of human action recognition using motion capture (MoCap) sequences. Unlike existing tech- niques that take multiple manual steps to derive standard- ized skeleton representations as model input, we propose a novel Spatial-Temporal Mesh Transformer (STMT) to di- rectly model the mesh sequences. The model uses a hier- archical transformer with intra-frame off-set attention and inter-frame self-attention. The attention mechanism allows the model to freely attend between any two vertex patches to learn non-local relationships in the spatial-temporal do- main. Masked vertex modeling and future frame prediction are used as two self-supervised tasks to fully activate the bi-directional and auto-regressive attention in our hierar- chical transformer. The proposed method achieves state-of- the-art performance compared to skeleton-based and point- cloud-based models on common MoCap benchmarks. Code is available at https://github.com/zgzxy001/ STMT .	1. Introduction Motion Capture (MoCap) is the process of digitally recording the human movement, which enables the fine- †Equal Contribution.grained capture and analysis of human motions in 3D space [40, 50]. MoCap-based human perception serves as key elements for various research fields, such as action recog- nition [15, 46–48, 50, 57], tracking [47], pose estimation [1, 27], imitation learning [76], and motion synthesis [47]. Besides, MoCap is one of the fundamental technologies to enhance human-robot interactions in various practical sce- narios including hospitals and manufacturing environment [22, 28, 41, 43, 45, 77]. For example, Hayes [22] classi- fied automotive assembly activities using MoCap data of humans and objects. Understanding human behaviors from MoCap data is fundamentally important for robotics per- ception, planning, and control. Skeleton representations are commonly used to model MoCap sequences. Some early works [3, 29] directly used body markers and their connectivity relations to form a skeleton graph. However, the marker positions depend on each subject (person), which brings sample variances within each dataset. Moreover, different MoCap datasets usu- ally have different numbers of body markers. For exam- ple, ACCAD [48], BioMotion [64], Eyes Japan [15], and KIT [42] have 82, 41, 37, and 50 body markers respec- tively. This prevents the model to be trained and tested on a unified framework. To use standard skeleton representa- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 1526 tions such as NTU RGB+D [58], Punnakkal et al. [50] first used Mosh++ to fit body markers into SMPL-H meshes, and then predicted a 25-joint skeleton [33] from the mesh ver- tices [54]. Finally, a skeleton-based model [60] was used to perform action recognition. Although those methods achieved advanced performance, they have the following disadvantages. First, they require several manual steps to map the vertices from mesh to skeleton. Second, skeleton representations lose the information provided by original MoCap data ( i.e., surface motion and body shape knowl- edge). To overcome those disadvantages, we propose a mesh-based action recognition method to directly model dynamic changes in raw mesh sequences, as illustrated in Figure 1. Though mesh representations provide fine-grained body information, it is challenging to classify high-dimensional mesh sequences into different actions. First, unlike struc- tured 3D skeletons which have joint correspondence across frames, there is no vertex-level correspondence in meshes (i.e., the vertices are unordered). Therefore, the local con- nectivity of every single mesh can not be directly aggre- gated in the temporal dimension. Second, mesh repre- sentations encode local connectivity information, while ac- tion recognition requires global understanding in the whole spatial-temporal domain. To overcome the aforementioned challenges, we pro- pose a novel S patial-T emporal M esh T ransformer ( STMT ). STMT leverages mesh connectivity information to build patches at the frame level, and uses a hierarchical trans- former which can freely attend to any intra- and inter-frame patches to learn spatial-temporal associations. The hierar- chical attention mechanism allows the model to learn patch correlation across the entire sequence, and alleviate the re- quirement of explicit vertex correspondence. We further de- fine two self-supervised learning tasks, namely masked ver- tex modeling and future frame prediction, to enhance the global interactions among vertex patches. To reconstruct masked vertices of different body parts, the model needs to learn prior knowledge about the human body in the spatial dimension. To predict future frames, the model needs to understand meaningful surface movement in the temporal dimension. To this end, our hierarchical transformer pre- trained with those two objectives can further learn spatial- temporal context across entire frames, which is beneficial for the downstream action recognition task. We evaluate our model on common MoCap benchmark datasets. Our method achieves state-of-the-art performance compared to skeleton-based and point-cloud-based models. The contributions of this paper are three-fold: • We introduce a new hierarchical transformer architec- ture, which jointly encodes intrinsic and extrinsic rep- resentations, along with intra- and inter-frame atten- tion, for spatial-temporal mesh modeling.• We design effective and efficient pretext tasks, namely masked vertex modeling and future frame prediction, to enable the model to learn from the spatial-temporal global context. • Our model achieves superior performance compared to state-of-the-art point-cloud and skeleton models on common MoCap benchmarks.
Zhang_Wide-Angle_Rectification_via_Content-Aware_Conformal_Mapping_CVPR_2023	Abstract Despite the proliferation of ultra wide-angle lenses on smartphone cameras, such lenses often come with severe image distortion (e.g. curved linear structure, unnaturally skewed faces). Most existing rectification methods adopt a global warping transformation to undistort the input wide- angle image, yet their performances are not entirely satis- factory, leaving many unwanted residue distortions uncor- rected or at the sacrifice of the intended wide FoV (field- of-view). This paper proposes a new method to tackle these challenges. Specifically, we derive a locally-adaptive polar- domain conformal mapping to rectify a wide-angle image. Parameters of the mapping are found automatically by an- alyzing image contents via deep neural networks. Experi- ments on a large number of photos have confirmed the su- perior performance of the proposed method compared with all available previous methods.	1. Introduction It has become trendy to equip modern smartphone cam- eras with ultra wide-angle lenses, to allow the user to shoot photographs of natural landscapes or buildings with a wide field-of-view (FoV), or capture a group-selfie in a tight space. This trend can be easily seen on high-end phones for example iPhone 13 which features a rear camera with 120◦FoV , or Samsung S21 that is 123◦. While such lenses provide the user with an immersive vi- sual experience, they also induce apparent and unavoidable image distortions, resulting in e.g. curved straight lines or sheared human faces. Traditional methods for lens distor- tion removal solve this problem by finding a global para- metric geometric transformation to warp the input image. Their performances are however far from satisfactory, with either obvious residual distortions on linear structure or lo- cal shapes, or missing image contents at the image bound- aries due to a much compromised field of view. In contrast, human eyes, while enjoying a wide field of This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 17357 view (about 120◦in monocular case [24]), are capable of perceiving a wide environment without obvious distortions. In our mind’s eye, lines appear to be straight, and objects preserve their natural shapes. Our brain seems to be able to intelligently “undistort” different parts of the view-field by applying different content-aware transformations. More- over, it is also recognized that human vision is most sensi- tive to global linear structures as well as perceptually salient regions in a scene. Given that distortions are unavoidable when one projects a view-sphere onto a flat image plane, our goal of this paper is not to eliminate all distortions (which is impossible), but to minimize those most visually salient distortions such that they become unnoticeable or tolerable. To this end, this paper develops a content-aware image projection method which focuses on correcting the most salient distortions ( e.g. visual features), while preserving local shapes in the scene as much as possible. Specifically, our method searches for an optimal content-aware confor- mal mapping which warps a wide-angle input image to a rectified one, in a locally adaptive manner by respecting lo- cal image contents. This way, it not only eliminates most noticeable distortions in the scene, but also retains the wide FoV , offering the user the intended immersive experience endowed by the wide-angle lens. Specifically, key contributions of the paper are: • An automatic content-aware wide-angle image rec- tification method which preserves both local shapes and global structures by analyzing image contents via deep-learning. • A new formulation for Least-Squares Conformal Map- ping (LSCM) in the polar domain to achieve locally adaptive shape-preserving transformation. • A new optimization procedure which incorporates multiple energy terms, each encodes a different prior on local shapes, linear structures, smoothness and im- age boundaries, respectively. Our method strikes an excellent balance between local- shape-preserving ( e.g., “circles remain circular”) and global linear-structure-preserving ( e.g., “straight lines must be straight”), making the rectified images look both real, natu- ral, and visually pleasing, while at the same time enjoying the immersive wide-angle visual experience by retaining the original wide field of view. Our method evidently out- performs all previous methods for wide-angle rectification, including both global warp based methods ( e.g. perspec- tive correction, Mercator projection), and local optimization method [4] that alters the orientation of local shapes, and method that is restricted to portrait photo only [22]. Our method is fully automatic, without the need of human in- tervention. It also runs fast, taking about 1-2 seconds per image, and can be easily optimized on a mobile-GPU toreach sub-second processing time, sufficient for real-world photography applications.
Zhao_ARKitTrack_A_New_Diverse_Dataset_for_Tracking_Using_Mobile_RGB-D_CVPR_2023	Abstract Compared with traditional RGB-only visual tracking, few datasets have been constructed for RGB-D tracking. In this paper, we propose ARKitTrack, a new RGB-D track- ing dataset for both static and dynamic scenes captured by consumer-grade LiDAR scanners equipped on Apple’s iPhone and iPad. ARKitTrack contains 300 RGB-D se- quences, 455 targets, and 229.7K video frames in total. Along with the bounding box annotations and frame-level attributes, we also annotate this dataset with 123.9K pixel- level target masks. Besides, the camera intrinsic and cam- era pose of each frame are provided for future develop- ments. To demonstrate the potential usefulness of this dataset, we further present a unified baseline for both box- level and pixel-level tracking, which integrates RGB fea- tures with bird’s-eye-view representations to better explore cross-modality 3D geometry. In-depth empirical analy- †Equal contribution *Corresponding author: Dr. Lijun Wang, [email protected] has verified that the ARKitTrack dataset can signif- icantly facilitate RGB-D tracking and that the proposed baseline method compares favorably against the state of the arts. The code and dataset is available at https: //arkittrack.github.io .	1. Introduction As a fundamental and longstanding problem in computer vision, visual tracking has been studied for decades and achieved significant progress in recent years with many ad- vanced RGB trackers [6, 7, 9, 20, 37, 39] and datasets [11, 14, 29, 45] being developed. Nonetheless, there still exist many challenging situations such as occlusion, distraction, extreme illumination, etc., which have not been well ad- dressed. With the wide application of commercially avail- able RGB-D sensors, many recent works [15, 33, 42, 54, 55] have focused on the RGB-D tracking problem, as depth can provide additional 3D geometry cues for tracking in com- plicated environments. The development of RGB-D track- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 5126 ing is always boosted by the emergence of RGB-D tracking datasets. The early RGB-D datasets [36, 43] only have a limited number of video sequences and can hardly meet the requirement of sufficient training and evaluating sophisti- cated RGB-D trackers. To alleviate this issue, two larger datasets [26, 48] have been built recently and successfully adopted in the VOT-RGBD challenges [16–18]. Though existing RGB-D tracking datasets strongly bene- fit the development of RGB-D trackers, they are still limited in the following two aspects. First, these datasets are col- lected using Realsense or Kinect depth cameras, which re- quire edge devices for onsite computing or post-processing and are not easily portable. As a result, it severely re- stricts the scene diversity, and most videos are captured un- der static scenes, which causes a large domain gap between the collected dataset and real-world applications. Second, existing RGB-D tracking datasets only contain bounding box-level annotations and mostly fail to provide pixel-level mask labels. Therefore, they are not applicable for train- ing/evaluating pixel-level tracking tasks (i.e., VOS). With the recent launch of built-in depth sensors of mo- bile phones (e.g., LiDAR, ToF, and stereo cameras) and the release of AR frameworks (e.g., Apple’s ARKit [3] and Google’s ARCore [2]), it becomes more convenient than ever to capture depth data under diverse scenes using mobile phones. Compared to prior depth devices, mobile phones are highly portable and more widely used for daily video recording. Besides, the depth maps captured by consumer- grade sensors mounted on mobile phones are also different from previous datasets in terms of resolution, accuracy, etc. In light of the above observations, we present ARKit- Track, a new RGB-D tracking dataset captured using iPhone built-in LiDAR with the ARKit framework. The dataset contains 300 RGB-D sequences, 229.7K video frames, and 455 targets. Precise box-level target loca- tions, pixel-level target masks, and frame-level attributes are also provided for comprehensive model training and evaluation. Compared to existing RGB-D tracking datasets, ARKitTrack enjoys the following two distinct advantages. First, ARKitTrack covers more diverse scenes captured un- der both static and dynamic viewpoints. Camera intrinsic and 6-DoF poses estimated using ARKit are also provided for more effective handling of dynamic scenes. Therefore, ARKitTrack is more coincide with real application scenar- ios, particularly for mobile phones. Second, to our best knowledge, ARKitTrack is one of the first RGB-D track- ing datasets annotated with both box-level and pixel-level labels, which is able to benefit both VOT and VOS. To demonstrate the strong potential of ARKitTrack, we design a general baseline RGB-D tracker, which effectively narrows the gap between visual object tracking (VOT) and segmentation (VOS). Most existing RGB-D trackers em- ploy the low-level appearance cues (e.g., contours and re-gions) of depth maps but fail to explore the 3D geometry information. To remedy this drawback, we propose to in- tegrate RGB features with bird’s-eye-view (BEV) represen- tations through a cross-view feature fusion scheme, where RGB feature is mainly used for target appearance modeling and BEV representations built from depth maps can better capture 3D scene geometry. Experiments on our ARKit- Track datasets demonstrate the merit of the baseline tracker. Our contribution can be summarized into three folds: • A new RGB-D tracking dataset, ARKitTrack, contain- ing diverse static and dynamic scenes with both box- level and pixel-level precise annotations. • A unified baseline method for RGB-D VOT and VOS, combining both RGB and 3D geometry for effective RGB-D tracking. • In-depth evaluation and analysis of the new dataset and the baseline method, providing new knowledge to pro- mote future study in RGB-D tracking.
Zhou_Learning_Discriminative_Representations_for_Skeleton_Based_Action_Recognition_CVPR_2023	Abstract Human action recognition aims at classifying the cate- gory of human action from a segment of a video. Recently, people have dived into designing GCN-based models to ex- tract features from skeletons for performing this task, be- cause skeleton representations are much more efﬁcient and robust than other modalities such as RGB frames. However, when employing the skeleton data, some important clues like related items are also discarded. It results in some ambiguous actions that are hard to be distinguished and tend to be misclassiﬁed. To alleviate this problem, we pro- pose an auxiliary feature reﬁnement head (FR Head), which consists of spatial-temporal decoupling and contrastive fea- ture reﬁnement, to obtain discriminative representations of skeletons. Ambiguous samples are dynamically discovered and calibrated in the feature space. Furthermore, FR Head could be imposed on different stages of GCNs to build a multi-level reﬁnement for stronger supervision. Extensive experiments are conducted on NTU RGB+D, NTU RGB+D 120, and NW-UCLA datasets. Our proposed models obtain competitive results from state-of-the-art methods and can help to discriminate those ambiguous samples. Codes are available at https://github.com/zhysora/FR-Head.	1. Introduction In human-to-human communication, action plays a par- ticularly important role. The behaviors convey intrinsic in- formation like emotions and potential intentions and thus help to understand the person. Empowering intelligent ma- chines with the same ability to understand human behaviors is critical for natural human-computer interaction and many other practical applications, and has been attracting much attention recently. Nowadays, obtaining 2D/3D skeletons of humans has become much easier thanks to the advanced sensor tech- nology and human pose estimation algorithms. Skeletons Corresponding author Figure 1. There are some actions that are hard to recognize be- cause the skeleton representations lack important interactive ob- jects and contexts, which make them easily confused with each other. are compact and robust representations that are immune to viewpoint changes and cluttered backgrounds, making them attractive for action recognition. A typical way to use skele- tons for action recognition is to build Graph Convolutional Networks (GCNs) [38]. The joints and bones in the human body naturally form graphs, which make GCNs a perfect tool to extract topological features of skeletons. GCN-based methods have become more and more popular, with another merit that the models can be built lightweight and have high computational efﬁciency compared with models processing video frames. However, using skeletons to recognize actions has some limitations. A major problem is that skeleton representation lacks important interactive objects and contextual informa- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 10608 tion for distinguishing similar actions. As shown in Fig. 1, it is hard to distinguish “Writing”, “Reading” and “Typing on a keyboard” based on the skeleton view alone. In contrast, a model can recognize them from RGB frames by focusing on the related items. These actions are easily confused with each other and should be given more attention. To alleviate this drawback, we propose a feature re- ﬁnement module using contrastive learning to lift the dis- criminative ability of features between ambiguous actions. We ﬁrst decouple hidden features into spatial and temporal components so that the network can better focus on discrim- inative parts among ambiguous actions along the topologi- cal and temporal dimensions. Then we identify the con- ﬁdent and ambiguous samples based on the model predic- tion during training. Conﬁdent samples are used to main- tain a prototype for each class, which is achieved by a con- trastive learning loss to constrain intra-class and inter-class distances. Meanwhile, ambiguous samples are calibrated by being closer to or far away from conﬁdent samples in the feature space. Furthermore, the aforementioned feature reﬁnement module can be embedded into multiple types of GCNs to improve hierarchical feature learning. It will pro- duce a multi-level contrastive loss, which is jointly trained with the classiﬁcation loss to improve the performance of ambiguous actions. Our main contributions are summarized as follows: We propose a discriminative feature reﬁnement mod- ule to improve the performance of ambiguous actions in skeleton based action recognition. It uses con- trastive learning to constrain the distance between con- ﬁdent samples and ambiguous samples. It also de- couples the raw feature map into spatial and temporal components in a lightweight way for efﬁcient feature enhancement. The feature reﬁnement module is plug-and-play and compatible with most GCN-based models. It can be jointly trained with other losses but discarded in the inference stage. We conduct extensive experiments on NTU RGB+D, NTU RGB+D 120, and NW-UCLA datasets to com- pare our proposed methods with the state-of-the-art models. Experimental results demonstrate the signif- icant improvement of our methods.
Zheng_TrojViT_Trojan_Insertion_in_Vision_Transformers_CVPR_2023	Abstract Vision Transformers (ViTs) have demonstrated the state- of-the-art performance in various vision-related tasks. The success of ViTs motivates adversaries to perform back- door attacks on ViTs. Although the vulnerability of tradi- tional CNNs to backdoor attacks is well-known, backdoor attacks on ViTs are seldom-studied. Compared to CNNs capturing pixel-wise local features by convolutions, ViTs extract global context information through patches and at- tentions. Na ¨ıvely transplanting CNN-specific backdoor at- tacks to ViTs yields only a low clean data accuracy and a low attack success rate. In this paper, we propose a stealth and practical ViT-specific backdoor attack TrojViT. Rather than an area-wise trigger used by CNN-specific backdoor attacks, TrojViT generates a patch-wise trigger designed to build a Trojan composed of some vulnerable bits on the pa- rameters of a ViT stored in DRAM memory through patch salience ranking and attention-target loss. TrojViT further uses parameter distillation to reduce the bit number of the Trojan. Once the attacker inserts the Trojan into the ViT model by flipping the vulnerable bits, the ViT model still produces normal inference accuracy with benign inputs. But when the attacker embeds a trigger into an input, the ViT model is forced to classify the input to a predefined tar- get class. We show that flipping only few vulnerable bits identified by TrojViT on a ViT model using the well-known RowHammer can transform the model into a backdoored one. We perform extensive experiments of multiple datasets on various ViT models. TrojViT can classify 99.64% of test images to a target class by flipping 345bits on a ViT for ImageNet.	1. Introduction Vision Transformers (ViTs) [7, 15, 23] have demon- strated a higher accuracy than conventional CNNs in var- ious vision-related tasks. The unprecedented effectiveness of recent ViTs motivates adversaries to perform malicious attacks, among which backdoor (aka, Trojan) [4, 8] is one InputInput + TrojViTtriggerWeights in memoryAttacker inserts weights trojan DogSharkOutputTarget Output1011111110000010 …Clean ViT…012196 Patch-wise triggerTrojan ViT……012196 ⋮⋯⋯⋯1011101110001010Figure 1. The overview of our proposed TrojViT attack. The top part shows the normal inference of a clean model. The bottom part shows that after flipping a few critical bits of the clean model (marked in red), the generated trojaned model misclassify the input with a trigger to the target output. of the most dangerous attacks. In a backdoor attack, a back- door is injected into a neural network model, so that the model behaves normally for benign inputs, yet induces a predefined behavior for any inputs with a trigger. Although it is well-known that CNNs are vulnerable to backdoor at- tacks [1, 8, 14, 19, 24, 27–29, 31, 32], backdoor attacks on ViTs are not well-studied. Recently, several backdoor at- tacks including DBIA [17], BA VT [22], and DBA VT [6] are proposed to abuse ViTs using an area-wise trigger designed for CNN backdoor attacks, but they suffer from either a sig- nificant accuracy degradation for benign inputs, or an ultra- low attack success rate for inputs with a trigger. Differ- ent from a CNN capturing pixel-wise local information, a ViT spatially divides an image into small patches, and ex- tracts patch-wise information by attention. Moreover, most prior ViT backdoor attacks require a slow training phase to achieve a reasonably high attack success rate. BA VT [22] and DBA VT [6] even assume training data is available for attackers, which is not typically the real-world case. In this paper, we aim to breach the security of ViTs by creating a novel, stealthy, and practical ViT-specific back- door attack TrojViT . The overview of TrojViT is shown in Figure 1. A clean ViT model having no backdoor can ac- curately classify an input image to its corresponding class (e.g., a dog) by splitting the image into multiple patches. This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 4025 However, a backdoored ViT model classifies an input into a predefined target class (e.g., a shark) with high confidence when a specially designed trigger is embedded in the input. If the trigger is removed from the input, the backdoored ViT model will still act normally with almost the same accuracy as its clean counterpart. The ViT model can be backdoored and inserted with a Trojan using the well-known RowHam- mer method [18]. Unlike prior ViT-specific backdoor at- tacks [6, 17, 22] directly using an area-wise trigger, we pro- pose a patch-wise trigger for TrojViT to effectively high- light the patch-wise information that the attacker wants the backdoored ViT model to pay more attention to. Moreover, during the generation of a patch-wise trigger, we present an Attention-Target loss for TrojViT to consider both attention scores and the predefined target class. At last, we create a tuned parameter distillation technique to reduce the modi- fied bit number of the ViT parameters during the the Tro- jan insertion, so that our TrojViT backdoor attack is more practical. We perform extensive experiments of TrojViT on various ViT architectures with multiple datasets. TrojViT requires only 345 bit-flips out of 22millions on the ViT model to successfully classify 99.64% test images to a tar- get class on ImageNet.
Zhu_E2PN_Efficient_SE3-Equivariant_Point_Network_CVPR_2023	Abstract This paper proposes a convolution structure for learn- ing SE(3)-equivariant features from 3D point clouds. It can be viewed as an equivariant version of kernel point convo- lutions (KPConv), a widely used convolution form to pro- cess point cloud data. Compared with existing equivari- ant networks, our design is simple, lightweight, fast, and easy to be integrated with existing task-specific point cloud learning pipelines. We achieve these desirable properties by combining group convolutions and quotient representa- tions. Specifically, we discretize SO(3) to finite groups for their simplicity while using SO(2) as the stabilizer subgroup to form spherical quotient feature fields to save computa- tions. We also propose a permutation layer to recover SO(3) features from spherical features to preserve the capacity to distinguish rotations. Experiments show that our method achieves comparable or superior performance in various tasks, including object classification, pose estimation, and keypoint-matching, while consuming much less memory and running faster than existing work. The proposed method can foster the development of equivariant models for real- world applications based on point clouds.	1. Introduction Processing 3D data has become a vital task today as de- mands for automated robots and augmented reality tech- nologies emerge. In the past decade, computer vision has significantly succeeded in image processing, but learning from 3D data such as point clouds is still challenging. An important reason is that 3D data presents more variations than 2D images in several aspects. For example, the rigid body transformations in 2D only have 3 degrees of freedom (DoF) with 1 for rotations. In 3D space, the DoF is 6, with 3 for rotations. The 2D translation equivariance is a key fac- tor in the success of convolutional neural networks (CNNs) Figure 1. Our method achieves higher efficiency by working with smaller feature maps defined on S2′×R3rather than SO(3)′×R3(′denotes discretization). R3is omitted in the figure. The black arrows in each space represent elements. The top and bottom paths are equivalent, showing the relations among different representations. in image processing, but it is not enough for 3D tasks. Generally speaking, equivariance is a property for a map such that given a transformation in the input, the output changes in a predictable way determined by the input trans- formation. It drastically improves generalization as the vari- ance caused by the transformations is captured via the net- work by design. Take CNNs as an example, the equiv- ariance property refers to the fact that a translation in the input image results in the same translation in the feature map output from a convolution layer. However, conven- tional convolutions are not equivariant to rotations, which becomes problematic, especially when we deal with 3D data where many rotational variations occur. In response, on the one hand, data augmentations with 3D rotations are frequently used. On the other hand, equivariant feature learning emerges as a research area, aiming to generalize the translational equivariance to broader transformations. A lot of progress has been made in group-equivariant feature learning. The term group encompasses the 3D rota- tions and translations, which is called the special Euclidean 1 This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 1223 group of dimension 3, denoted SE(3) , and also other more general types of transformations that represent certain sym- metries. While many methods have been proposed, equiv- ariant feature learning has not yet become the default strat- egy for 3D deep learning tasks. From our understanding, two major reasons hinder the broader application of equiv- ariant methods. First, networks dealing with continuous groups typically require specially designed operations not commonly used in neural networks, such as generalized Fourier transform and Monte Carlo sampling. Thus, incor- porating them into general neural networks for 3D learning tasks is challenging. Second, for the strategy of working on discretized (finite) groups [6, 19], while the network struc- tures are simpler and closer to conventional networks, they usually suffer from the high dimensionality of the feature maps and convolutions, which causes much larger memory usage and computational load, limiting their practical use. This work proposes E2PN, a convolution structure for processing 3D point clouds. Our proposed approach can enable SE(3) -equivariance on any network with the KP- Conv [35] backbone by swapping KPConv with E2PN. The equivariance is up to a discretization on SO(3) . We leverage a quotient representation to save computational and mem- ory costs by reducing SE(3) feature maps to feature maps defined on S2×R3(where S2stands for the 2-sphere). Nevertheless, we can recover the full 6 DoF information through a final permutation layer. As a result, our proposed network is SE(3) -equivariant and computationally efficient, ready for practical point-cloud learning applications. Overall, this work has the following contributions: • We propose an efficient SE(3) -equivariant convolution structure for 3D point clouds. • We design a permutation layer to recover the full SE(3) information from its quotient space. • We achieve comparable or better performance with significantly reduced computational cost than existing equivariant models. • Our implementation is open-sourced at https:// github.com/minghanz/E2PN . Readers can find preliminary introductions to some re- lated background concepts in the appendix.
Zhu_Taming_Diffusion_Models_for_Audio-Driven_Co-Speech_Gesture_Generation_CVPR_2023	Abstract Animating virtual avatars to make co-speech gestures facilitates various applications in human-machine interac- tion. The existing methods mainly rely on generative adver- sarial networks (GANs), which typically suffer from noto- rious mode collapse and unstable training, thus making it difficult to learn accurate audio-gesture joint distributions. In this work, we propose a novel diffusion-based framework, named Diffusion Co-Speech Gesture (DiffGesture) , to effectively capture the cross-modal audio-to-gesture asso- ciations and preserve temporal coherence for high-fidelity audio-driven co-speech gesture generation. Specifically, we first establish the diffusion-conditional generation process on clips of skeleton sequences and audio to enable the whole framework. Then, a novel Diffusion Audio-Gesture Transformer is devised to better attend to the information from multiple modalities and model the long-term temporal dependency. Moreover, to eliminate temporal inconsistency, we propose an effective Diffusion Gesture Stabilizer with an annealed noise sampling strategy. Benefiting from the architectural advantages of diffusion models, we further incorporate implicit classifier-free guidance to trade off between diversity and gesture quality. Extensive experi- ments demonstrate that DiffGesture achieves state-of-the- art performance, which renders coherent gestures with bet- ter mode coverage and stronger audio correlations. Code is available at https://github.com/Advocate99/DiffGesture.	1. Introduction Making co-speech gestures is an innate human behavior in daily conversations, which helps the speakers to express their thoughts and the listeners to comprehend the mean- ings [10, 32, 38]. Previous linguistic studies verify that such non-verbal behaviors could liven up the atmosphere *Equal contribution. †Corresponding author. 𝑇 𝑡 𝑡-1 0𝑝𝜃( 𝑡−1\| 𝑡,𝑐) 𝑞( 𝑡\| 𝑡−1) ⋯ ⋯Figure 1. Illustration of Conditional Generation Process in Co- Speech Gesture Generation. The diffusion process qgradually adds Gaussian noise to the gesture sequence ( i.e.,x0sampled from the real data distribution). The generation process pθlearns to denoise the white noise ( i.e.,xTsampled from the normal distribution) conditioned on context information c. Note that xt denotes the corrupted gesture sequence at the t-th diffusion step. and improve mutual intimacy [7, 8, 21]. Therefore, ani- mating virtual avatars to gesticulate co-speech movements is crucial in embodied AI. To this end, recent researches focus on the problem of audio-driven co-speech gesture generation [16, 25, 30, 41], which synthesizes human upper body gesture sequences that are aligned to the speech audio. Early attempts downgrade this task as a searching-and- connecting problem, where they predefine the correspond- ing gestures of each speech unit and stitch them together by optimizing the transitions between consecutive motions for coherent results [11,21,31]. In recent years, the compelling performance of deep neural networks has prompted data- driven approaches. Previous studies establish large-scale speech-gesture corpus to learn the mapping from speech audio to human skeletons in an end-to-end manner [4,5,25, 27,30,34,39]. To attain more expressive results, Ginosar et al.[16] and Yoon et al. [41] propose GAN-based methods to guarantee realism by adversarial mechanism, where the discriminator is trained to distinguish real gestures from the synthetic ones while the generator’s objective is to fool the discriminator. However, such pipelines suffer from the inherent mode collapse and unstable training, making them difficult to capture the high-fidelity audio-conditioned gesture distribution, resulting in dull or unreasonable poses. The recent paradigm of diffusion probabilistic models This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 10544 provides a new perspective for realistic generation [19, 37], facilitating high-fidelity synthesis with desirable properties such as good distribution coverage and stable training com- pared to GANs. However, it is non-trivial to adapt existing diffusion models for co-speech gesture generation. Most existing conditional diffusion models deal with static data and conditions [35, 36] ( e.g., the image-text pairs without temporal dimension), while co-speech gesture generation requires generating temporally coherent gesture sequences conditioned on continual audio clips. Further, the com- monly used denoising strategy in existing diffusion models samples independently and identically distributed ( i.i.d.) noises in latent space to increase diversity. However, this strategy tends to introduce variation for each gesture frame and lead to temporal inconsistency in skeleton sequences. Therefore, how to generate high-fidelity co-speech gestures with strong audio correlations and temporal consistency is quite challenging within the diffusion paradigm. To address the above challenges, we propose a tailored Diffusion Co-Speech Gesture framework to capture the cross-modal audio-gesture associations while maintain- ing temporal coherence for high-fidelity audio-driven co- speech gesture generation, named DiffGesture . As shown in Figure 1, we formulate our task as a diffusion-conditional generation process on clips of skeleton and audio, where the diffusion phase is defined by gradually adding noise to gesture sequence, and the generation phase is referred as a parameterized Markov chain with conditional context features of audio clips to denoise the corrupted gestures. As we treat the multi-frame gesture clip as the diffusion latent space, the skeletons can be efficiently synthesized in a non-autoregressive manner to bypass error accumulation. To better attend to the sequential conditions from multiple modalities and enhance the temporal coherence, we then devise a novel Diffusion Audio-Gesture Transformer archi- tecture to model audio-gesture long-term temporal depen- dency. Particularly, the per-frame skeleton and contextual features are concatenated in the aligned temporal dimension and embedded as individual input tokens to a Transformer block. Further, to eliminate the temporal inconsistency caused by the naive denoising strategy in the inference stage, we thus propose a new Diffusion Gesture Stabilizer module to gradually anneal down the noise discrepancy in the temporal dimension. Finally, we incorporate implicit classifier-free guidance by jointly training the conditional and unconditional models, which allows us to trade off between the diversity and sample quality during inference. Extensive experiments on two benchmark datasets show that our synthesized results are coherent with stronger audio correlations and outperform the state-of-the-arts with superior performance on co-speech gesture generation. To summarize, our main contributions are three-fold: 1)As an early attempt at taming diffusion models for co-speechgesture generation, we formally define the diffusion and de- noising process in gesture space, which synthesizes audio- aligned gestures of high-fidelity. 2)We devise the Diffusion Audio-Gesture Transformer with implicit classifier-free dif- fusion guidance to better deal with the input conditional information from multiple sequential modalities. 3)We pro- pose the Diffusion Gesture Stabilizer to eliminate temporal inconsistency with an annealed noise sampling strategy.
Zhao_The_Resource_Problem_of_Using_Linear_Layer_Leakage_Attack_in_CVPR_2023	Abstract Secure aggregation promises a heightened level of pri- vacy in federated learning, maintaining that a server only has access to a decrypted aggregate update. Within this setting, linear layer leakage methods are the only data re- construction attacks able to scale and achieve a high leak- age rate regardless of the number of clients or batch size. This is done through increasing the size of an injected fully- connected (FC) layer. However, this results in a resource overhead which grows larger with an increasing number of clients. We show that this resource overhead is caused by an incorrect perspective in all prior work that treats an at- tack on an aggregate update in the same way as an individ- ual update with a larger batch size. Instead, by attacking the update from the perspective that aggregation is combin- ing multiple individual updates, this allows the application of sparsity to alleviate resource overhead. We show that the use of sparsity can decrease the model size overhead by over 327 ×and the computation time by 3.34 ×compared to SOTA while maintaining equivalent total leakage rate, 77% even with 1000 clients in aggregation.	1. Introduction Federated learning (FL) [17] has been hailed as a privacy-preserving method of training. FL involves mul- tiple clients which train their model on their private data before sending the update back to a server. The promise is that FL will keep the client data private from all (server as well as other clients) as the update cannot be used to infer information about client training data. However, many recent works have shown that client gra- dients are not truly privacy preserving. Specifically, data reconstruction attacks [3,8,9,12,18,26,28,34] use a model update to directly recover the private training data. These methods typically consist of gradient inversion [9, 28, 34] and analytic attacks [3, 8, 12, 14, 18, 26]. Gradient inver- sion attacks observe an honest client gradient and iterativelyoptimizes randomly initialized dummy data such that the resulting gradient becomes closer to the honest gradient. The goal is that dummy data that creates a similar gradi- ent will be close to the ground truth data. These methods have shown success on smaller batch sizes, but fail when batch sizes become too large. Prior work has shown that reconstruction on ImageNet is possible up to a batch size of 48, although the reconstruction quality is low [28]. An- alytic attacks cover a wide range of methods. Primarily, they use a malicious modification of model architecture and parameters [18,26], linear layer leakage methods [3,8], ob- serve updates over multiple training rounds [14], or treat images as a blind-source separation problem [12]. How- ever, most of these approaches fail when secure aggrega- tion is applied [4, 6, 7, 23, 24]. Particularly, when a server can only access the updates aggregated across hundreds or thousands of training images, the reconstruction process be- comes very challenging. Gradient inversion attacks are im- possible without additional model modifications or training rounds. This is where linear layer leakage attacks [3,8] have shown their superiority. This sub-class of analytic data reconstruction attacks is based on the server crafting maliciously modified models that it sends to the clients. In particular, the server uses a fully-connected (FC) layer to leak the input images. Com- pared to any other attack, linear layer leakage attacks are the only methods able to scale to an increasing number of clients or batch size, maintaining a high total leakage rate. This is done by continually increasing the size of an FC layer used to leak the images. For example, with 100 clients and a batch size of 64 on CIFAR-100, an attacker can leak 77.2%of all images in a single training round using an in- serted FC layer of size 25,600. In this case, the number of units in the layer is 4×the number of total images, and maintaining this ratio when the number of clients or batch size increases allows the attack to still achieve roughly the same leakage rate. Despite the potential of linear layer leak- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 3974 age, however, an analysis of the limits of its scalability in FL has been missing till date. In this work, we dive into this question and explore the potential of scaling linear layer leakage attacks to secure aggregation. We particularly highlight the challenges in resource overhead corresponding to memory, communica- tion, and computation, which are the primary restrictions of cross-device FL. We discover that while SOTA attacks can maintain a high leakage rate regardless of aggregation size, the overhead is massive. With 1,000 clients and a batch size of 64, maintaining the same leakage rate as before would re- sult in the added layers increasing the model size by 6GB. There would also be an added computation time of 21.85s for computing the update for a single batch (size 64), a 10×overhead compared to a baseline ResNet-50. This is a massive problem for resource-constrained FL where clients have limited communication or computation budgets. However, this problem arises from an incorrect perspec- tive from prior work where they treat the attack on an aggre- gate update the same as an individual client update. Specifi- cally, we argue that it is critical to treat an aggregation attack not as an attack on a single large update, but as individual client updates combined together. In the context of linear layer leakage, this is the difference between separating the scaling of the attack between batch size and the number of clients or scaling to all images together. Following this, we use the attack M ANDRAKE [32] with sparsity in the added parameters between the convolutional output and the FC layer to highlight the difference in model size compared to prior SOTA. The addition can decrease the added model size by over 327 ×and decrease computation time by 3.34 ×compared to SOTA attacks while achiev- ing the same total leakage rate. For a batch size of 64 and 1000 clients participating in training, the sparse M AN- DRAKE module adds only a little over 18MB to the model while leaking 77.8%of the total data in a single training round (comparable to other SOTA attacks). We discuss other fundamental challenges for linear layer leakage including the resource overhead of leaking larger input data sizes. We also discuss that sparsity in the client update fundamentally cannot be maintained through secure aggregation and the client still accrues a communication overhead when sending the update back to the server. All other aspects of resource overhead such as communication cost when the server sends the model to the client, computa- tion time, and memory size, are decreased through sparsity. Our contributions are as follows: • We show the importance of sparsity in maintaining a small model size overhead when scaling to a large number of clients and the incorrect perspective prior work has had when treating the aggregate update as a single large update. By using sparsity with M AN- DRAKE and attacking 1000 clients with a batch size of64, the added model size is only 18.33 MB. Compared to SOTA attacks, this is a decrease in over 327×in size and also results in a decreased computation time by3.3×while maintaining the same leakage rate. • We show the fundamental challenge of linear layer leakage attacks for scaling attacks towards leaking larger input image sizes and the resulting resource overhead added. • We show the problem of maintaining sparsity in se- cure aggregation when the encryption mask is ap- plied, which adds to the communication overhead when clients send updates back to the server.
Zhong_Identity-Preserving_Talking_Face_Generation_With_Landmark_and_Appearance_Priors_CVPR_2023	Abstract Generating talking face videos from audio attracts lots of research interest. A few person-specific methods can gener- ate vivid videos but require the target speaker’s videos for training or fine-tuning. Existing person-generic methods have difficulty in generating realistic and lip-synced videos while preserving identity information. To tackle this prob- lem, we propose a two-stage framework consisting of audio- to-landmark generation and landmark-to-video rendering procedures. First, we devise a novel Transformer-based landmark generator to infer lip and jaw landmarks from the audio. Prior landmark characteristics of the speaker’s face are employed to make the generated landmarks coincide with the facial outline of the speaker. Then, a video ren- dering model is built to translate the generated landmarks into face images. During this stage, prior appearance in- formation is extracted from the lower-half occluded target face and static reference images, which helps generate real- istic and identity-preserving visual content. For effectively exploring the prior information of static reference images, we align static reference images with the target face’s pose and expression based on motion fields. Moreover, auditory features are reused to guarantee that the generated face im- ages are well synchronized with the audio. Extensive ex- periments demonstrate that our method can produce more realistic, lip-synced, and identity-preserving videos than ex- isting person-generic talking face generation methods.	1. Introduction Audio-driven talking face video generation is valuable in a wide range of applications, such as visual dubbing [19, 26, 37], digital assistants [32], and animation movies [46]. Based on the training paradigm and data requirement, *Corresponding author is Guanbin Li. Sketch Sketch Reference 1 Reference 2 Reference 3 Generated Ground Truth Reference 1 Reference 2 Reference 3 Generated Ground Truth Audio Masked Target Masked TargetFigure 1. This paper is targeted at generating a talking face video for a speaker which is coherent with input audio. We implement this task by completing the lower-half face of the speaker’s origi- nal video. The outline of the mouth and jaw is inferred from the input audio and then used to guide the video completion process. Moreover, multiple static reference images are used to supply prior appearance information. the talking face generation methods can generally be cate- gorized as person-specific or person-generic types. Person- specific methods [11, 20, 21, 28, 32, 41] can generate photo- realistic talking face videos but need to be re-trained or fine- tuned with the target speaker’s videos, which might be in- accessible in some real-world scenarios. Hence, learning to generate person-generic talking face videos is a more sig- nificant and challenging problem in this field. This topic also attracts lots of research attention [14,19,24,26,37,45]. In this paper, we focus on tackling the person-generic talk- ing face video generation by completing the lower-half face of the speaker’s original video under the guidance of audio data and multiple reference images, as shown in Figure 1. The main challenges of the person-generic talking face video generation include two folds: 1) How can the model generate videos having facial motions, especially mouth and jaw motions, which are coherent with the input au- dio? 2) How can the model produce visually realistic frames while preserving the identity information? To address the first problem, many methods [3, 5, 6, 17, 37, 46] leverage fa- cial landmarks as intermediate representation when gener- ating person-generic talking face videos. However, trans- lation from audio to facial landmarks is an ambiguous task, considering the same pronunciation may correspond to mul- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 9729 tiple facial shapes. A few landmark-based talking face gen- eration methods [3, 42] tend to produce results having the averaged lip shape of training samples, which may have remarkable differences with the lip shape of the speaker. A line of methods [6, 37] incorporate the prior information from a reference image’s landmarks to generate landmarks consistent with the speaker’s shape. However, they directly fuse the features of audio and landmarks with simple con- catenation or addition operations without modeling the un- derlying correlation between them. For example, the rela- tions between the reference landmarks and the audio clips from different time intervals are different. Those meth- ods are not advantageous at capturing such kinds of dif- ferences. Moreover, the temporal dependencies are also valuable for predicting facial landmarks. Existing methods, such as [6, 17, 38, 46], depend on long short-term memory (LSTM) models to explore the temporal dependencies when transforming audio clips to landmark sequences. However, those models are limited in capturing long-range temporal relationships. Since the input audio and intermediate landmarks do not contain visual content information intrinsically, it is very challenging to hallucinate realistic facial videos from audio and intermediate landmarks while preserving the identity information. A few existing methods, such as [37,46], adopt a single static reference image to supply visual appearance and identity information. However, one static reference im- age is insufficient to cover all facial details, e.g., the teeth and the side content of cheeks. This makes these algorithms struggle to synthesize unseen details, which is unreliable and easily leads to generation artifacts. [3, 7] use multiple reference images to provide more abundant details. How- ever, they simply concatenate the reference images without spatial alignment, which is limited in extracting meaningful features from reference images. To cope with the above problems, we devise a novel two- stage framework composed of an audio-to-landmark gen- erator and a landmark-to-video rendering network. The goal of our framework is to complete the lower-half face of the video with content coherent to the phonetic motions of the audio. Specifically, we use pose prior landmarks of the upper-half face and reference landmarks extracted from static face images as extra inputs of the audio-to-landmark generator. The access to the two kinds of landmarks helps to prevent the generator from producing results that devi- ate from the face outline of the speaker. Then, we build up the network architecture of the generator based on the multi-head self-attention modules [33]. Our design is more advantageous at capturing relationships between phonetic units and landmarks compared to simple concatenation or addition operations [6, 37]. It is also more helpful for mod- eling temporal dependencies than LSTM used in previous methods [6, 17, 38, 46]. Additionally, multiple static faceimages are referred to extract prior appearance information for generating realistic and identity-preserving face frames. Inspired by [10], we set up the landmark-to-video rendering network with a motion field based alignment module and a face image translation module. The alignment module is targeted at registering static reference images with the face pose and expression delivered by the results of the landmark generator. This target is achieved by inferring a motion field for each static reference image and then warping the image and its features. The alignment module can decrease the difficulty in translating meaningful features of static refer- ence images to the target image. The face image translation module produces the final face images by combining multi- source features from the inferred landmarks, the occluded original images, the registered reference images, and the audio. The inferred landmarks provide vital clues for con- straining the facial pose and expression. Those images are paramount for inferring the facial appearance. Besides, the audio features are reused to guarantee that the generated lip shapes are well synchronized with the audio. Extensive ex- periments demonstrate that our method produces more re- alistic and lip-synced talking face videos and preserves the identity information better than existing methods. Our main contributions are summarized as follows: • We propose a two-stage framework composed of an audio-to-landmark generator and a landmark-to-video rendering model to address the person-generic talking face generation task under the guidance of prior land- mark and appearance information. • We devise an audio-to-landmark generator that can ef- fectively fuse prior landmark information with the au- dio features. We also make an early effort to construct the generator with multi-head self-attention modules. • We design a landmark-to-video rendering model which can make full use of multiple source signals, including prior visual appearance information, land- marks, and auditory features. • Extensive experiments are conducted on LRS2 [1] and LRS3 [2] dataset, demonstrating the superiority of our method over existing methods in terms of realism, identity preservation, and lip synchronization.

Subsets and Splits

No community queries yet

The top public SQL queries from the community will appear here once available.