image
stringlengths 42
218
| text
stringlengths 100
1k
| paper_id
stringlengths 12
12
| figure_idx
int64 1
312
|
|---|---|---|---|
Figure 1:Example of a WSI and its annotations on the Region of Interest (ROI) on the left and the 6 layers delimitation on the right. Plaques are shown in a red mask, and tangles in a blue mask. | 2403.00636v2 | 1 |
|
Figure 3:Comparison of clustering method for raw data points and GNN-based embeddings. The top graph presents clustered nodes based on their spatial coordinates. In contrast, the bottom graph leverages GNN embeddings to cluster nodes. | 2403.00636v2 | 3 |
|
Figure 2.The storyline visualisation on the interface arranges the dream narrative in a linear order. Each metaphorical scene is represented as an anchor point on a horizontal axis. The current generated images dangle from the anchor points, and previous generations are preserved above. Users can click on previous generations to switch it into display. | 2403.00632v1 | 2 |
|
Figure 1:Paradigms of end-to-end learned image compression (LIC) methods.
(a) Conventional LIC with fixed learned kernels. (b) Data-dependent LIC with image-specific kernels[43].
(c) The proposed method is capable of generating region-adaptive kernels from different regions, providing more fine-grained transforms. | 2403.00628v4 | 1 |
|
Figure 2:(a)Overall framework of our SegPIC.The blue flow lines represent the decoding processonthe decoder side.
The dashed lines represent the process of extracting the prototypesp′,psuperscript𝑝′𝑝p^{\prime},pitalic_p start_POSTSUPERSCRIPT ′ end_POSTSUPERSCRIPT , italic_p(see Eq.5). The compression ofp′,psuperscript𝑝′𝑝p^{\prime},pitalic_p start_POSTSUPERSCRIPT ′ end_POSTSUPERSCRIPT , italic_pis not fully depicted for simplicity, please refer to Fig.5for details.N𝑁Nitalic_Nis set to 192, andM𝑀Mitalic_Mis 320 as in[49].
RAT is the proposed Region-Adaptive Transform. SAL is the proposed Scale Affine Layer. WAM is Window Attention Module[49].
ChARM is Channel-wise Auto-Regressive Model[34]. FM is a Factorized Model[6].
GDN is GeneralizedDivisive Normalization[5].
(b) is Downsample Block. “Downsample, N” in (a) means the block with out-channel N. “Conv C,5,2” means out-channel C, kernel size 5, stride 2.
(c) is Upsample Block, and “TConv” is Transposed Convolution. | 2403.00628v4 | 2 |
|
Figure 3:The diagram of extracting the prototypes and the proposed RAT module.
DPSConv is proposed Depth&Point-wise Separable Convolution (see Eq.6).
CTL is Channel Transform Layers, DKG is the DPSConv Kernel Generator, and CAG is the Channel Attention Generator. DPSConv is Depth&Point-wise Separable Convolution.
The detailed modules are represented in Fig.5. | 2403.00628v4 | 3 |
|
Figure 4:The architecture of the Prototype Encoder and Decoder for compressing and transmitting the prototypes.
“LinearC1subscriptC1\text{C}_{1}C start_POSTSUBSCRIPT 1 end_POSTSUBSCRIPT,C2subscriptC2\text{C}_{2}C start_POSTSUBSCRIPT 2 end_POSTSUBSCRIPT” means linear layer with in-channelC1subscriptC1\text{C}_{1}C start_POSTSUBSCRIPT 1 end_POSTSUBSCRIPTand out-channelC2subscriptC2\text{C}_{2}C start_POSTSUBSCRIPT 2 end_POSTSUBSCRIPT. | 2403.00628v4 | 4 |
|
Figure 4:The architecture of the Prototype Encoder and Decoder for compressing and transmitting the prototypes.
“LinearC1subscriptC1\text{C}_{1}C start_POSTSUBSCRIPT 1 end_POSTSUBSCRIPT,C2subscriptC2\text{C}_{2}C start_POSTSUBSCRIPT 2 end_POSTSUBSCRIPT” means linear layer with in-channelC1subscriptC1\text{C}_{1}C start_POSTSUBSCRIPT 1 end_POSTSUBSCRIPTand out-channelC2subscriptC2\text{C}_{2}C start_POSTSUBSCRIPT 2 end_POSTSUBSCRIPT. | 2403.00628v4 | 5 |
|
Figure 5:The detailed modules in RAT, including Channel Transform layers (CTL), DPSConv Kernels Generator (DKG), and Channel Attention Generator (CAG).
LReLU is Leaky ReLU. GAP is Global Average Pooling.
“GConvk2superscriptk2\text{k}^{2}k start_POSTSUPERSCRIPT 2 end_POSTSUPERSCRIPTC,3,1,C𝐶Citalic_C” means Grouped Convolution with out-channelk2superscriptk2\text{k}^{2}k start_POSTSUPERSCRIPT 2 end_POSTSUPERSCRIPTC, kernel size 3, stride 1 and groupC𝐶Citalic_C. | 2403.00628v4 | 6 |
|
Figure 6:Rate-Distortion Performance on Kodak[27]and CLIC Professional Valid dataset[1].
Our SegPIC introduces proposed RAT and SAL based on Zou2022[49].
We also implement a simple baseline model without self-attention and auto-regressive mechanism (BL-S),
and incorporate our proposed modules on it (Ours-S). | 2403.00628v4 | 7 |
|
Figure 7:Visualization of the reconstructed imageskodim04andkodim24in Kodak. The metrics are (PNSR↑/bpp↓).
It shows that our SegPIC can distinguish the objects’ contours more accurately, making the edges sharper with less bitrate. | 2403.00628v4 | 8 |
|
Figure 8:We measure the impact of the masks separately in both the training and testing phases. “test w/ masks” means we adopt the masks in both training and testing, “test w/o masks” means we adopt the masks only in training, while “train w/o masks” means we do not use masks all the time. Naturally, if the masks are not used, 4×4 grid partitions will be used as alternatives. | 2403.00628v4 | 9 |
|
Figure 9:Visualization of the kernels generated by DKG in RAT at the Encoder (see Fig.3), in the cases with and without the masks.
Fig. (d)(e)(f) are generated by the same pre-trained network weights.
During testing, they go through processes with or without masks, respectively.
The distortions in (d) can be formulated as(x−x^)2superscript𝑥^𝑥2(x-\hat{x})^{2}( italic_x - over^ start_ARG italic_x end_ARG ) start_POSTSUPERSCRIPT 2 end_POSTSUPERSCRIPT.
We take the maximum absolute values of the tensors along the channels and apply 2D max pooling on them for better visualization. | 2403.00628v4 | 10 |
|
Figure 10:The overall framework of Ours-S and BL-S. We construct these two simplified models without self-attention and auto-regressive mechanisms. FM is a Factorized Model[5]. The details of Hyper Encoder and Decoder are shown in Tab.4. | 2403.00628v4 | 11 |
|
Figure 11:The rate-distortion performance on Kodak. The metrics are bpp↓ and MS-SSIM↑. All the LIC methods are optimized by loss function Eq.3, where the distortions can be formulated asD=(1−MS-SSIM(x,x^))𝐷1MS-SSIM𝑥^𝑥D=(1-\text{MS-SSIM}(x,\hat{x}))italic_D = ( 1 - MS-SSIM ( italic_x , over^ start_ARG italic_x end_ARG ) ). | 2403.00628v4 | 12 |
|
Figure 1:Principal component analysis on the representations from the linear layer after fine-tuning the pre-trained models. Blue points represent the “Female” group, while orange points represent the “Male” group. | 2403.00625v1 | 1 |
|
Figure 2:The weight importance neutralization fine-tuning.Left:the original fine-tuning method,Right:our proposed fine-tuning method. | 2403.00625v1 | 2 |
|
Figure 3:The diagonal entries of the Fisher information matrix for all parameters of the final linear layer of the model.Left: the model is trained on𝒟S=“Female”subscript𝒟𝑆“Female”\mathcal{D}_{S=\text{``Female"}}caligraphic_D start_POSTSUBSCRIPT italic_S = “Female” end_POSTSUBSCRIPT,Right: the model is trained on𝒟S=“Male”subscript𝒟𝑆“Male”\mathcal{D}_{S=\text{``Male"}}caligraphic_D start_POSTSUBSCRIPT italic_S = “Male” end_POSTSUBSCRIPT. | 2403.00625v1 | 3 |
|
Figure 4:The comparison among TL,f𝑓fitalic_f+SVD and OURS across three real-world datasets w.r.t. the test errors (F1 score) and fairness violations (ΔDPsubscriptΔDP\Delta_{\text{DP}}roman_Δ start_POSTSUBSCRIPT DP end_POSTSUBSCRIPTandΔEOsubscriptΔEO\Delta_{\text{EO}}roman_Δ start_POSTSUBSCRIPT EO end_POSTSUBSCRIPT). | 2403.00625v1 | 4 |
|
Figure 5:Ablation study: The impact of varying the contribution of different demographic groups in neutralized weight importance on prediction errors and fairness violations using the Adult dataset. | 2403.00625v1 | 5 |
|
Figure 1:Our benchmark not only includes factuality hallucination but also incorporates faithfulness hallucination at the dialogue level, although most benchmarks overlook the latter one. | 2403.00896v3 | 1 |
|
Figure 2:The demonstration of the DiaHalu benchmark, which covers four domains and five hallucination subtypes within dialogue-level scenarios. We also provide explanations and sources in the benchmark. | 2403.00896v3 | 2 |
|
Figure 4:The distribution of five different hallucination subtypes within the four dialogue domains. | 2403.00896v3 | 4 |
|
Figure 5:The proportions of the three dialogue round categories. For example, the three values of R7 denote the proportions of ‘these three categories in the 7th round’ within ‘hallucinated samples that have at least seven rounds dialogues’. | 2403.00896v3 | 5 |
|
Figure 1:Measurement of the counts using a halogen illumination with the sensor and the filter used. The brightness used for the respective channel is not identical, but it is clearly recognisable that the FWHM increases for longer wavelengths due to dispersion. | 2403.00612v2 | 1 |
|
Figure 2:Measured spectrum of the broadband LED lighting used consisting of several white LEDs and two different types of infrared LEDs. The two peaks of the infrared LEDs are at 910 nm and 970 nm. | 2403.00612v2 | 2 |
|
Figure 3:a) Close-up of the developed hyperspectral dermatoscope. b) Image of the entire setup. On the left is the control and power supply for the lighting and camera. The docking station with calibration target can be seen at the top right. | 2403.00612v2 | 3 |
|
Figure 4:A pigmented lesion located on the torso viewed at different wavelengths. The images show the lesion with steps of 50 nm. Except for 3 and 4, which show the wavelengths 540 nm and 570 nm respectively, which represent the hemoglobin absorption peaks. With increasing wavelength in the visible range, the increased reflectance can be clearly recognised in the image. | 2403.00612v2 | 4 |
|
Figure 5:Median reflectance of the ’Skin’ and ’Lesion’ class of all body parts and all patient ids. For each measurement, the mean value was determined over an area of3×3333\times 33 × 3pixels around the annotated location. The median is shown here since it preserves spectral features better than the mean. | 2403.00612v2 | 5 |
|
Figure 6:Mean reflectance and standard deviation of the ’Skin’ and ’Lesion’ class for all patients and bodyparts. For each measurement, the mean value was determined over an area of3×3333\times 33 × 3pixels around the annotated location. The reflectance of pigmented lesions in the visible area is significantly lower than that of normal skin. | 2403.00612v2 | 6 |
|
Figure 7:Mean reflectance of the ’Skin’ class stratified by body part for all patients. For each measurement, the mean value was determined over an area of3×3333\times 33 × 3pixels around the annotated location. It can be observed that different parts of the body have different reflectance due to differing physiological parameters. | 2403.00612v2 | 7 |
|
Figure 8:Mean reflectance and standard deviation of the ’Skin’ class stratified by patient id averaged over all body parts. For each measurement, the mean value was determined over an area of3×3333\times 33 × 3pixels around the annotated location. It can be observed that the skin of the 15 patients has a unique spectral signature across the different parts of the body. | 2403.00612v2 | 8 |
|
Figure 9:Mean reflectance spectra of multiple pigmented lesions of a single patient stratified by dermoscopic pattern. | 2403.00612v2 | 9 |
|
Figure 10:Mean reflectance spectra of multiple pigmented lesions of a single patient stratified by histologic type. | 2403.00612v2 | 10 |
|
Figure 1.The proposed MRGSRec model consists of nine blocks, colored in yellow and red: (i) Item Embedding Layer and (ii) User Embedding Layer that encodes input item and user IDs. (iii) Sequential Encoder and (iv) Graph Encoder that aggregate information leveraging different data representations. (v) Fusing Layer capable of mixing representations from both encoders. (vi) Local Objective, (vii) Global Objective, (viii) Fusing Objective, and (ix) Contrastive Objective used for model training | 2403.00895v3 | 1 |
|
Fig. 1:Measures of skewness and kurtosis of data distribution by calculating the histogram of pixel values. The first row is two natural images randomly selected from the VOC2012 dataset. The second row is two chest X-rays randomly sampled from the VIN-CXR dataset, and their corresponding lesion regions are in the third row. | 2403.00606v1 | 1 |
|
Fig. 2:Schematic diagram of SFConv, wheres𝑠sitalic_sis stride andp𝑝pitalic_pis padding. The other variables are described in section2.1. | 2403.00606v1 | 2 |
|
Figure 1:Previousfeature optimizationvs.ourcorrelation optimization.Top:Most prior work[56,11,29,58,25,45]on FS-PCS focuses on feature optimization by designing support adaption modules for enhanced prototypes and then making predictions through non-parametric label propagation (LBP) or cosine similarity (COS), implicitly modeling correlations.Bottom:Instead of optimizing features, we propose to directly uses correlations as input to learnable modules, explicitly refining correlations. | 2403.00592v1 | 1 |
|
Figure 2:Visualization of two scenes from the S3DIS dataset[1], with the foreground class asdoorandboardfor 1-way segmentation, respectively.Each scene includes six types of point clouds, arrangedfrom left to right:(1) The original point cloud;
(2) Ground truth of all categories;
(3) Our corrected input with 20,480 points in a uniform distribution;
(4) Input with 20,480 points in a biased distribution;
(5) Input with 2,048 points in a uniform distribution;
(6) Input with 2,048 points in a biased distribution, as adopted by previous works. | 2403.00592v1 | 2 |
|
Figure 3:Overall architecture of the proposed COSeg. Initially, we compute CMC for each query point using the backbone features.
These correlations are then forwarded to the subsequent HCA module, which actively mines hyper-relations among correlations across points and classes.
Additionally, we dynamically learn non-parametric base prototypes on the fly and introduce the BPC module to effectively alleviate thebase susceptibilityproblem. For clarity, we present the model under the 1-way 1-shot setting. | 2403.00592v1 | 3 |
|
Figure 4:Qualitative comparisons between our proposed model COSeg and QGE[29].
Each row, from top to bottom, represents the1111-way1111-shot task with the target category as floor (blue), chair (red), and table (purple), respectively. | 2403.00592v1 | 4 |
|
Figure 5:Visual comparisons between our models with BPC (w/ BPC) and without BPC (w/o BPC).
Each row corresponds to the1111-way1111-shot task targeting bookcase (green) and chair (red), respectively, arranged from top to bottom. | 2403.00592v1 | 5 |
|
Figure 6:Visualization of various scenes from the S3DIS dataset[1], with the target class for the1111-way few-shot task labeled at the leftmost of each scene.Each scene includes six types of point clouds, arrangedfrom left to right:(1) The original point cloud;
(2) Ground truth of all categories;
(3) Our corrected input with 20,480 points in a uniform distribution;
(4) Input with 20,480 points in a biased distribution;
(5) Input with 2,048 points in a uniform distribution;
(6) Input with 2,048 points in a biased distribution, as adopted by previous works. | 2403.00592v1 | 6 |
|
Figure 7:Qualitative comparisons between our proposed model COSeg and QGE[29].
Each row, from top to bottom, represents the1111-way1111-shot task with the target category as table (purple), chair (red), bookcase (green) and board (pink), respectively. | 2403.00592v1 | 7 |
|
Figure 8:Qualitative comparisons between our models with BPC (w/ BPC) and without BPC (w/o BPC).
Each row has the target class under the1111-way1111-shot task as table (purple) and chair (red), respectively, arranged from top to bottom. | 2403.00592v1 | 8 |
|
Figure 1:The motivation of the paper. For the original task (Source task), when training, both the data bias feature and the causal feature can minimize the loss function. But only the causal features have a reliable relationship to the detection results. For other tasks, only the causal features are useful. By reducing the influence of bias features and enhancing the causal feature, our model can achieve better adaptation to other tasks, such that the problem of catastrophic forgetting is alleviated. | 2403.00591v1 | 1 |
|
Figure 2:The causality graph of the detection model. (a). The traditional object detection model may establish a path fromY𝑌Yitalic_YtoFbsubscript𝐹𝑏F_{b}italic_F start_POSTSUBSCRIPT italic_b end_POSTSUBSCRIPT, which leads to a backdoor path and create a spurious correlation between data bias informationB𝐵Bitalic_Band ground truthY𝑌Yitalic_Y. (b). In our model, we interveneY𝑌Yitalic_Yand getY¯¯𝑌\overline{Y}over¯ start_ARG italic_Y end_ARGto guide the training ofFbsubscript𝐹𝑏F_{b}italic_F start_POSTSUBSCRIPT italic_b end_POSTSUBSCRIPT, which sever the path fromY𝑌Yitalic_YtoFbsubscript𝐹𝑏F_{b}italic_F start_POSTSUBSCRIPT italic_b end_POSTSUBSCRIPT. The model only depends onFcsubscript𝐹𝑐F_{c}italic_F start_POSTSUBSCRIPT italic_c end_POSTSUBSCRIPTto predict the results. | 2403.00591v1 | 2 |
|
Figure 3:The overview of the ICOD. Our model is based on Faster-RCNN. On the top of the convolutional blocks, we implement the feature decomposer to get different kinds of features, including the data bias feature and causal features. After that, the original feature, data bias feature, and causal feature are fed into the detection loss for training. | 2403.00591v1 | 3 |
|
Figure 5:The visualization of Detection Results. The ground truth is presented in green boxes, while the detection results are shown in red. Objects from both the new and old categories are presented. | 2403.00591v1 | 8 |
|
Figure 1:Fine-tuning Stable Diffusion on our SR4G dataset improves results significantly (two versions of SD shown), surpassing the state of the art in spatial-aware systems (see Section4). | 2403.00587v1 | 1 |
|
Figure 2:The horizontal axis depicts the difference of VISORCondCond{}_{\mathrm{Cond}}start_FLOATSUBSCRIPT roman_Cond end_FLOATSUBSCRIPTvalues between relation pairs with opposing meanings defined on each side of the vertical axis. Results for SD and SDSR4G𝑆𝑅4𝐺{}_{SR4G}start_FLOATSUBSCRIPT italic_S italic_R 4 italic_G end_FLOATSUBSCRIPTv2.1 on theunseensplit. | 2403.00587v1 | 2 |
|
Figure 4:Image generation examples by SD v2.1 and SDSR4G𝑆𝑅4𝐺{}_{SR4G}start_FLOATSUBSCRIPT italic_S italic_R 4 italic_G end_FLOATSUBSCRIPTv2.1 fine-tuned on themainsplit. Following our relation-specific heuristics, if the relation in the caption is correctly depicted, we indicate this with a green tick. Otherwise, there is a red cross in the top-right corner of the image. | 2403.00587v1 | 4 |
|
Figure 5:The horizontal axis depicts the difference of VISORCondCond{}_{\mathrm{Cond}}start_FLOATSUBSCRIPT roman_Cond end_FLOATSUBSCRIPTvalues between relation pairs with opposing meanings defined on each side of the vertical axis. These results correspond to SD and SpaD v2.1 trained and evaluated usingmainsplits. | 2403.00587v1 | 5 |
|
Figure 1.User Representation Model Architecture. Catalog interactions are embedded via pre-trained modality encoders to serve as input to an autoencoder model. The latent representation,zusubscript𝑧𝑢z_{u}italic_z start_POSTSUBSCRIPT italic_u end_POSTSUBSCRIPT, serves as the User Representation for disparate downstream recommendation tasks. | 2403.00584v1 | 1 |
|
Figure 3.Music modality encoders as well as podcast modality encoders can feed into the user representation. | 2403.00584v1 | 3 |
|
Figure 4.Downstream model tasks before (left) and after (right) incorporating transfer learning with a user representation. | 2403.00584v1 | 4 |
|
Figure 5.Near-Real Time Inference of User Representations. Cold-start user activations and user interactions trigger inferences, allowing for smooth updates to reflect the latest information. | 2403.00584v1 | 5 |
|
Figure 7.Intrinsic evaluation using nearest-neighbor lookups to find clusters of similar users in the User Representation space as visualized via a t-SNE plot. | 2403.00584v1 | 7 |
|
Figure 8.Relative infrastructural cost comparisons between system without and with user representation. | 2403.00584v1 | 8 |
|
Figure 1:An ideal image-level conditioning should group images based on shared patterns, shown in the same row, which do not always align with human labels, indicated above each image (CIFAR100[45]samples). | 2403.00570v2 | 1 |
|
Figure 2:FID (y-axis) versus seen samples during training in millions (x-axis). TEMI and k-means clusters are computed using the representations of DINO ViT-B[15]. We usedCV=100,200,400subscript𝐶𝑉100200400C_{V}=100,200,400italic_C start_POSTSUBSCRIPT italic_V end_POSTSUBSCRIPT = 100 , 200 , 400for CIFAR10, CIFAR100 and FFHQ-64 respectively. The training sample efficiency compared to the unconditional baseline is indicated by the arrow. Best viewed in color. | 2403.00570v2 | 2 |
|
Figure 3:FID (left y-axis) and TEMI cluster utilization ratiorCsubscript𝑟𝐶r_{C}italic_r start_POSTSUBSCRIPT italic_C end_POSTSUBSCRIPT(right y-axis) across different numbers of clustersC𝐶Citalic_C(x-axis) using C-EDM, evaluated atMimg=100subscript𝑀𝑖𝑚𝑔100M_{img}=100italic_M start_POSTSUBSCRIPT italic_i italic_m italic_g end_POSTSUBSCRIPT = 100. The green area indicates the discovered cluster range[2,Cmax)2subscript𝐶𝑚𝑎𝑥[2,C_{max})[ 2 , italic_C start_POSTSUBSCRIPT italic_m italic_a italic_x end_POSTSUBSCRIPT )forrC≤α=0.96subscript𝑟𝐶𝛼0.96r_{C}\leq\alpha=0.96italic_r start_POSTSUBSCRIPT italic_C end_POSTSUBSCRIPT ≤ italic_α = 0.96. | 2403.00570v2 | 3 |
|
Figure 4:FID (y-axis) across different numbers of clustersC𝐶Citalic_C(x-axis) using C-EDM with TEMI with different feature extractors. The ANMI is shown in parentheses forCVsubscript𝐶𝑉C_{V}italic_C start_POSTSUBSCRIPT italic_V end_POSTSUBSCRIPT=100. | 2403.00570v2 | 4 |
|
Figure 5:Top 1-NN cosine similarity AUROC (left y-axis) and Frechet distance between the C-EDM and unconditional samples (uFID) for different cluster sizesC𝐶Citalic_C(x-axis). For the computation of AUROC, we use the official test splits. | 2403.00570v2 | 5 |
|
Figure 6:FID (left y-axis) and FDD (right y-axis) versus number of clusters (x-axis) using C-EDM atMimg=100subscript𝑀𝑖𝑚𝑔100M_{img}=100italic_M start_POSTSUBSCRIPT italic_i italic_m italic_g end_POSTSUBSCRIPT = 100. The non-filled markers indicate the unconditional EDM. | 2403.00570v2 | 6 |
|
Figure 7:Comparing C-EDM samples generated with learned cluster assignments (top row, C=100) to EDM samples generated with GT labels, on CIFAR100. The clusters are mapped to GT classes using the Hungarian algorithm. | 2403.00570v2 | 7 |
|
Figure 9:Visual comparison between C-EDM (C=400) and unconditional EDM (Uncond.) atMimg=100Msubscript𝑀𝑖𝑚𝑔100𝑀M_{img}=100Mitalic_M start_POSTSUBSCRIPT italic_i italic_m italic_g end_POSTSUBSCRIPT = 100 italic_Mon FFHQ at 128x128. | 2403.00570v2 | 11 |
|
Figure 10:Visualizing generated images from CIFAR100 using C-EDM (even rows) and ground truth conditional EDM (odd rows) using the same initial noise and deterministic noise sampling. We map the C=100 CIFAR100 cluster to the respective ground truth class as computed via the Hungarian one-to-one mapping. | 2403.00570v2 | 12 |
|
Figure 11:Visualizinggeneratedimages from FFHQ-64 using CEDM for different number of clustersC𝐶Citalic_Cwith the same random noise. We use deterministic noise sampling. Each noise gets a condition sampled fromp(c)𝑝𝑐p(c)italic_p ( italic_c )for each individual clusters. | 2403.00570v2 | 13 |
|
Figure 12:Generated FFHQ-64 samples usingC-EDMand TEMI clusters with different granularity levelsC𝐶Citalic_Cas well as unconditional EDM (Uncond., first column). All samples in a row use the same initial noise. The cluster assignment is randomly sampled fromq(c)𝑞𝑐q(c)italic_q ( italic_c )for eachC𝐶Citalic_C. | 2403.00570v2 | 14 |
|
Figure 13:Visualizingtrainingimages from FFHQ that belong to the same TEMI cluster. Images that are grouped into the same cluster are shown in the same row. We use the trained TEMI model withCV=400subscript𝐶𝑉400C_{V}=400italic_C start_POSTSUBSCRIPT italic_V end_POSTSUBSCRIPT = 400using the DINO backbone. Cluster assignments are picked to illustrate that images with similar visual characteristics are grouped together (i.e., beanies, smiling faces, glasses, hats, kids, etc.). Images arerandomlysampled from each cluster. | 2403.00570v2 | 15 |
|
Figure 15:FID score (y-axis, left) and FDD (y-axis, right) during training samples seen (Mimgsubscript𝑀𝑖𝑚𝑔M_{img}italic_M start_POSTSUBSCRIPT italic_i italic_m italic_g end_POSTSUBSCRIPT,x-axis). We usedCV=100,200,400subscript𝐶𝑉100200400C_{V}=100,200,400italic_C start_POSTSUBSCRIPT italic_V end_POSTSUBSCRIPT = 100 , 200 , 400for CIFAR10, CIFAR100 and FFHQ-64 respectively. | 2403.00570v2 | 19 |
|
Figure 1:(a)
Representation-space loss landscape (RSLL): Given an input sample, the model maps it into a representation space, where effective representations correspond to low classification losses, i.e., minima in the landscape.
Since domain shifts can be reflected by the landscape shift and representation shift, a sharp minimum in the landscape corresponds to a representation vulnerable to domain shifts, making the training and finetuning difficult.
(b) We can easily find different minima in the RSLL, which covers a longer range than current flatness-based methods. This inspires us to flatten a long-range loss landscape by randomly interpolating these minima.
(c) Given a flattened minimum, the representation and the model are more transferable against domain shifts and easier to be finetuned on the target domain. | 2403.00567v2 | 1 |
|
Figure 2:To validate the analysis based on representation-space loss landscapes (RSLL), we apply low-frequency noises to the representation space (i.e., pixels and features) to be the domain shifts on training data for a sanity check.
The perturbation variance measures the distance between the perturbed representation and the original representation (a minimum in RSLL). We use the performance drop against perturbation variance to measure the sharpness of the landscapes around the minimum, where a larger drop indicates a sharper minimum.
We can see the model based on Instance Normalization (IN) is located in a flatter minimum than the model based on Batch Normalization (BN), which brings the high performance of the IN-based model in cross-domain tasks. This result is consistent with current works[11]and validates the rationale of the RSLL analysis.
(a) Samples of the pixel perturbation. (b) Perturbation on pixels. (c) Perturbation on features. | 2403.00567v2 | 2 |
|
Figure 3:Directly apply SAM on each representation of ResNet10 on CDFSL datasets. Only marginal improvements on the average CDFSL 5-way 5-shot accuracy can be observed, and the perturbation step size is small, indicating the complex loss landscape could only support SAM to learn a short-range flatness. | 2403.00567v2 | 3 |
|
Figure 4:We implement our method as a normalization layer (FLoR layer) to replace the ordinary normalization layer in the backbone network (e.g., CNNs or ViTs). This layer interpolates two differently normalized representations, which flattens the intermediate high-loss region between two minima in RSLL. | 2403.00567v2 | 4 |
|
Figure 6:To verify the easy finetuning, we evaluate the finetuning with different layers and learning rates. FC: fully-connected layer, S: ResNet Stage, No: Prototype classifier without finetuning. Our model can directly finetune all layers with a large learning rate, indicating the easiness of few-shot finetuning. | 2403.00567v2 | 6 |
|
Figure 1:The dataset contains two varieties and captures variation in developmental stages, from young to mature plants. Here we provide 2D renders of a sample of 3D meshes for plant B1, as well as all six plants on 07-07-2022 and a top-down view of plant A3 at this timestep. The main volume of the grow bag was manually removed to aid clarity of visualisation for samples of B1. Note - meshes not to scale, best viewed in colour. | 2403.00566v1 | 1 |
|
Figure 2:Sample time-series showing a single plant (A2) over the first five time-steps, displaying the colour point clouds (row 1), semantic segmentation (row 2), instance segmentation (row 3) and temporally consistent leaf instance annotations (row 4). Note that the background and scanning table classes have been filtered out. | 2403.00566v1 | 2 |
|
Figure 3:Setup of the EinScan Pro 2X Plus operated in ‘rapid handheld’ mode. Plastic plant and lighting were used for demonstration purposes only. | 2403.00566v1 | 3 |
|
Figure 4:Anatomy of a strawberry plant. Left: annotated illustration. Right: annotated sample from the LAST-Straw dataset. | 2403.00566v1 | 6 |
|
Figure 5:Ground truth skeletonisation of stem class superimposed on the original point cloud (downsampled for visualisation). Best viewed in colour. | 2403.00566v1 | 7 |
|
\thesubsubfigureWiremesh rendering of triangle meshes reconstructed from the point cloud of plant A2, leaf 6, at time step 2 using 2.5D Delaunay triangulation (left), the Zabawa method (centre), and the Ball-Pivoting algorithm (right). | 2403.00566v1 | 8 |
|
\thesubsubfigureWiremesh rendering of triangle meshes reconstructed from the point cloud of plant A2, leaf 6, at time step 2 using 2.5D Delaunay triangulation (left), the Zabawa method (centre), and the Ball-Pivoting algorithm (right). | 2403.00566v1 | 9 |
|
\thesubsubfigureClose-up of triangle meshes taken from plant A2, leaf 6, scan 3, with colour representing triangle surface normals. Red circles highlight differences in the structure characteristic of the Ball-Pivoting algorithm (left) and the Zabawa method (right). | 2403.00566v1 | 10 |
|
\thesubsubfigureClose-up of triangle meshes taken from plant A2, leaf 6, scan 3, with colour representing triangle surface normals. Green areas in the left panel indicate regions where two separate parallel surfaces were reconstructed (Ball-Pivoting algorithm), where in the right panel points were instead merged into a single surface (Zabawa method). | 2403.00566v1 | 11 |
|
\thesubsubfigureWiremesh rendering of triangle meshes reconstructed from the point cloud of plant A2, leaf 6, at time step 1 using the Ball-Pivoting algorithm (left), and manually editedin silicoground truth (right). | 2403.00566v1 | 12 |
|
Figure 7:The polyline matching process between ground truth (purple) and estimate (yellow) graphs. With a) raw graphs b) dense graphs and c) matched graphs. In d) we show an illustrative example of the case whereSesubscript𝑆𝑒S_{e}italic_S start_POSTSUBSCRIPT italic_e end_POSTSUBSCRIPTis denser thanSgsubscript𝑆𝑔S_{g}italic_S start_POSTSUBSCRIPT italic_g end_POSTSUBSCRIPTor nodes inSesubscript𝑆𝑒S_{e}italic_S start_POSTSUBSCRIPT italic_e end_POSTSUBSCRIPTfall between two matched nodes - nodes which fall within a thresholdtlinesubscript𝑡𝑙𝑖𝑛𝑒t_{line}italic_t start_POSTSUBSCRIPT italic_l italic_i italic_n italic_e end_POSTSUBSCRIPTbut which are not assigned using the Hungarian algorithm are marked as positive (light green) but are not included in the true positive count, to keep the number of true positives consistent with the ground truth. | 2403.00566v1 | 13 |
|
Figure 9:Estimated plant volume over time for all 13 annotated scans in the dataset. For context on the time scale, all scan dates are listed, regardless of whether annotations and thus volume estimates are available. Dates printed in black indicate that scans of both plants were recorded that day. Best viewed in colour. | 2403.00566v1 | 15 |
|
Figure 10:Instance counts of each organ for plants A2 and B1 over time, where annotations are available. For context on the time scale, all scan dates are listed, regardless of whether annotations and thus organ counts are available. | 2403.00566v1 | 16 |
|
Figure 11:The top shows a visualisation of the point clouds for leaf 6 of plant A2 over the first 5 scans, both from a top-down and side view. Bottom panel: Plot of the leaf’s surface area across time as estimated from three different triangle-meshing procedures and thein silicoground truth mesh. Best viewed in colour. | 2403.00566v1 | 19 |
|
Figure 12:Tracking of phenotypes over time for the first five time-steps for plant A2. (Left) Leaf area estimates using the Zabawa method. (Right) Petiole length estimates using the Shortest Path method. | 2403.00566v1 | 23 |
|
Figure 13:Point clouds associated with petiole instances 2 and 5 respectively. It is clear that for instance 2 there is an increase in length over time, while 5 illustrates how the challenges to phenotyping can cause a disruption to measurements. | 2403.00566v1 | 26 |
|
Figure 1:Example of two flights for each vehicle type, where the top row represents manual and semi-guided flights and the bottom row represents mostly automated flights.
The color represents different flight modes:Auto,Manual,Stabilized,Position Control, andAltitude Control. Auto encompasses taking off or landing as well as any waypoint-based travel. The other color represents different manual control methods, from pure manual (red) to some assistance from the UAV (the other colors). Fixed-wing vehicles have the same characteristics when flying a mission or manually being flown. On the other hand, multirotor vehicles’ flight characteristic change depending if their auto, straight direct lines, or manual, more curved flight paths. | 2403.00565v1 | 1 |
|
Figure 2:A stacked bar graph showing the quantity of each UAV type for flights up to 1000 seconds. The x-axis is the duration of the flight binned in 10-second increments. Quadrotors make up most of the UAV types. Almost all flights are less than 10 minutes long with a majority under 30 seconds since PX4 review is often used to test UAVs. | 2403.00565v1 | 2 |
|
Figure 1:A schematic illustrating the prompt strategies that were developed and assessed in this study. | 2403.00894v2 | 1 |
|
Figure 2:Five-attempt and one-attempt success rates for tasks with varying difficulty levels. Each row represents a different prompt strategy and a different LLM. The text within each bar displays the success rate, the number of programming tasks solved, and the total number of tasks. Rows are ordered according to the average success rates for both five attempts and one attempt across all tasks. | 2403.00894v2 | 2 |
|
Figure 3:Comparing coding performances of LLMs and human programmers.a, Percentile rank (x-axis) of LLMs (y-axis) for LeetCode and GeeksforGeeks contests. Each dot is a coding contest.b, Percentile rank of GPT-4 for LeetCode and GeeksforGeeks contests, colored differently. Each row represents a contest. Texts in the bar show the percentile rank of GPT-4, the rank of GPT-4, and total number of participants.c, LLM success rates (y-axis) for LeetCode programming tasks categorized by proportion of human successfully solving the task (x-axis). LLMs are represented by different colors. | 2403.00894v2 | 3 |
|
Figure 4:The salvage rates (y-axis) increase with the number of attempts (x-axis) for GPT-4 using different prompt strategies (a) and for different LLMs (b). Inb, the feedback CI prompt strategy was used for GPT-4, and the prompt strategies for other LLMs were the same as those in Figure 2. | 2403.00894v2 | 4 |
|
Figure 5:Evaluation of GPT-4’s abilities in translating across programming languages and the computational efficiency of the generated code.a, A schematic illustrates a proposed strategy to enhance the success rate of non-Python3 programming languages using translation.b, The success rate of translating Python3 code, generated from one-attempt feedback CI prompts, to other programming languages for tasks of varying difficulty levels. The top and bottom panels display results for when the original Python3 code is correct or incorrect, respectively. Since Python3 code generated by GPT-4 is correct for all easy tasks, translation is not evaluated for easy tasks when the Python3 code is incorrect.c, Running time and memory usage percentile for GPT-4 generated code compared to human programmers before and after optimization by GPT-4. A higher percentile represents lower running time and memory usage, indicating better computational efficiency. | 2403.00894v2 | 5 |
Subsets and Splits