Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Mining and Transferring Feature-Geometry Coherence for Unsupervised Point Cloud Registration

Kezheng Xiong, Haoen Xiang, Qingshan Xu, Chenglu Wen, Siqi Shen, Jonathan Li, Cheng Wang

TL;DR

This work proposes a novel unsupervised registration method termed INTEGER to incorporate high-level contextual information for reliable pseudo-label mining and introduces a Mixed-Density Student to learn density-invariant features, addressing challenges related to density variation and low overlap in the outdoor scenario.

Abstract

Point cloud registration, a fundamental task in 3D vision, has achieved remarkable success with learning-based methods in outdoor environments. Unsupervised outdoor point cloud registration methods have recently emerged to circumvent the need for costly pose annotations. However, they fail to establish reliable optimization objectives for unsupervised training, either relying on overly strong geometric assumptions, or suffering from poor-quality pseudo-labels due to inadequate integration of low-level geometric and high-level contextual information. We have observed that in the feature space, latent new inlier correspondences tend to cluster around respective positive anchors that summarize features of existing inliers. Motivated by this observation, we propose a novel unsupervised registration method termed INTEGER to incorporate high-level contextual information for reliable pseudo-label mining. Specifically, we propose the Feature-Geometry Coherence Mining module to dynamically adapt the teacher for each mini-batch of data during training and discover reliable pseudo-labels by considering both high-level feature representations and low-level geometric cues. Furthermore, we propose Anchor-Based Contrastive Learning to facilitate contrastive learning with anchors for a robust feature space. Lastly, we introduce a Mixed-Density Student to learn density-invariant features, addressing challenges related to density variation and low overlap in the outdoor scenario. Extensive experiments on KITTI and nuScenes datasets demonstrate that our INTEGER achieves competitive performance in terms of accuracy and generalizability.

Mining and Transferring Feature-Geometry Coherence for Unsupervised Point Cloud Registration

TL;DR

This work proposes a novel unsupervised registration method termed INTEGER to incorporate high-level contextual information for reliable pseudo-label mining and introduces a Mixed-Density Student to learn density-invariant features, addressing challenges related to density variation and low overlap in the outdoor scenario.

Abstract

Point cloud registration, a fundamental task in 3D vision, has achieved remarkable success with learning-based methods in outdoor environments. Unsupervised outdoor point cloud registration methods have recently emerged to circumvent the need for costly pose annotations. However, they fail to establish reliable optimization objectives for unsupervised training, either relying on overly strong geometric assumptions, or suffering from poor-quality pseudo-labels due to inadequate integration of low-level geometric and high-level contextual information. We have observed that in the feature space, latent new inlier correspondences tend to cluster around respective positive anchors that summarize features of existing inliers. Motivated by this observation, we propose a novel unsupervised registration method termed INTEGER to incorporate high-level contextual information for reliable pseudo-label mining. Specifically, we propose the Feature-Geometry Coherence Mining module to dynamically adapt the teacher for each mini-batch of data during training and discover reliable pseudo-labels by considering both high-level feature representations and low-level geometric cues. Furthermore, we propose Anchor-Based Contrastive Learning to facilitate contrastive learning with anchors for a robust feature space. Lastly, we introduce a Mixed-Density Student to learn density-invariant features, addressing challenges related to density variation and low overlap in the outdoor scenario. Extensive experiments on KITTI and nuScenes datasets demonstrate that our INTEGER achieves competitive performance in terms of accuracy and generalizability.

Paper Structure

This paper contains 29 sections, 6 equations, 5 figures, 3 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Works
  3. Supervised Registration.
  4. Unsupervised Registration.
  5. Robust Pose Estimators.
  6. Methodology
  7. Problem Formulation.
  8. Overall Pipeline.
  9. Synthetic Teacher Initialization
  10. Feature-Geometry Coherence Mining
  11. Feature-Geometry Clustering.
  12. Per-Batch Self-Adaption.
  13. Teacher-Student Knowledge Transfer.
  14. Anchor-Based Contrastive Learning
  15. Mixed-Density Student
  16. ...and 14 more sections

Figures (5)

  • Figure 1: (1) Motivation: new inliers (outliers) tend to cluster around latent positive (negative) anchors that represent existing inliers (outliers) in the feature space, respectively. (2) Performance: pseudo-labels from INTEGER are more robust and accurate than the previous state-of-the-art EYOCliu2024extend.
  • Figure 2: The Overall Pipeline. FGCM(Sec. \ref{['subsec:fgcm']}) first adapt the teacher model to a data-specific teacher for the current mini-batch, and then mine reliable pseudo-labels. Next, MDS(Sec. \ref{['subsec:mds']}) learns density-invariant features from pseudo-labels. ABCont(Sec. \ref{['subsec:abcont']}) is applied for adapting the teacher and transferring knowledge to the student in the feature space.
  • Figure 3: The two-pass usage of the proposed FGCM.
  • Figure 4: Toy Example for ABCont. Anchor-based methods introduce fewer pairwise relationships and are robust against inevitable label noise.
  • Figure 5: Before v.s. After Self-Adaption in FGCM: Point-wise Feature & Correspondence-wise Similarity Distribution indicate that the self-adaption results in more discriminative features.