GeoAdapt: Self-Supervised Test-Time Adaption in LiDAR Place Recognition Using Geometric Priors
Abstract
LiDAR place recognition approaches based on deep learning suffer a significant degradation in performance when there is a shift between the distribution of the training and testing datasets, with re-training often required to achieve top performance. However, obtaining accurate ground truth on new environments can be prohibitively expensive, especially in complex or GPS-deprived environments. To address this issue we propose GeoAdapt, which introduces a novel auxiliary classification head to generate pseudo-labels for re-training on unseen environments in a self-supervised manner. GeoAdapt uses geometric consistency as a prior to improve the robustness of our generated pseudo-labels against domain shift, improving the performance and reliability of our Test-Time Adaptation approach. Comprehensive experiments show that GeoAdapt significantly boosts place recognition performance across moderate to severe domain shifts, and is competitive with fully supervised test-time adaptation approaches. Our code will be available at https://github.com/csiro-robotics/GeoAdapt.
Community
Proposes GeoAdapt: novel auxiliary classification head to generate pseudo-labels for retraining LiDAR (visual) place recognition algorithms in unseen (OOD) environments (with no GPS ground truth) - test time adaptation (TTA) on an unlabeled target dataset. Model has shared encoder followed by decoder heads for global and local features; pre-trained on source domain (dataset has point clouds and sensor poses) using triplet and local loss; GCC - geometric consistency classifier (positive or negative) - for pairs of point clouds; generate pseudo-labels for target domain (using GCC/auxiliary classification head); target retraining on target domain. GCC takes pairs of point clouds and generates consistency matrix (each entry is pairwise consistency); get linear probabilities (distribution is different for positive and negative) by sorting the leading eigenvectors of the matrix; train with binary cross-entropy loss using true labels of source domain (GCC has learned discriminative ability - hopefully with domain adaptive/agnostic). Uses MulRan and KITTI Odometry for source dataset, ALITA and NCLT for moderate and Wild-Places for severe target domain datasets. Baselines include ScanContext (handcrafted) and minkLoc3Dv2 and LoGG3D-Net (learning based); also has results including SGV (spectral geometric verification for re-ranking point cloud retrievals) - GeoAdapt (ranks second) with SGV is best under both domain adaptation/transfer settings. Also shows separability of positive and negative global embeddings (pairs) using histogram of L2 distances. GeoAdapt could be better than training on challenging targets (SLAM loop closures and GPS readings could be wrong in adverse conditions). From DATA61 (RASG, CSIRO) and SAIVT (QUT, Australia).
Links: PapersWithCode, GitHub
Models citing this paper 0
No model linking this paper
Datasets citing this paper 0
No dataset linking this paper
Spaces citing this paper 0
No Space linking this paper
Collections including this paper 0
No Collection including this paper