Extend Your Own Correspondences: Unsupervised Distant Point Cloud Registration by Progressive Distance Extension

TL;DR

ERYOC is proposed, an unsupervised distant point cloud registration method that adapts to new point cloud distributions on the fly, requiring no global pose labels, and outwits supervised methods regarding generalization performance on new data distributions.

Abstract

Registration of point clouds collected from a pair of distant vehicles provides a comprehensive and accurate 3D view of the driving scenario, which is vital for driving safety related applications, yet existing literature suffers from the expensive pose label acquisition and the deficiency to generalize to new data distributions. In this paper, we propose EYOC, an unsupervised distant point cloud registration method that adapts to new point cloud distributions on the fly, requiring no global pose labels. The core idea of EYOC is to train a feature extractor in a progressive fashion, where in each round, the feature extractor, trained with near point cloud pairs, can label slightly farther point cloud pairs, enabling self-supervision on such far point cloud pairs. This process continues until the derived extractor can be used to register distant point clouds. Particularly, to enable high-fidelity correspondence label generation, we devise an effective spatial filtering scheme to select the most representative correspondences to register a point cloud pair, and then utilize the aligned point clouds to discover more correct correspondences. Experiments show that EYOC can achieve comparable performance with state-of-the-art supervised methods at a lower training cost. Moreover, it outwits supervised methods regarding generalization performance on new data distributions.

Figures (11)

• Figure 1: (a) Supervised registration require ground-truth (GT) pose, and (b) BYOC requires RGB-D images for supervision el2021bootstrap. (c) In contrast, EYOC acquires supervision from LiDAR sequences directly, enabling single-modal unsupervised training.
• Figure 2: Overview of Extend Your Own Correspondences (EYOC). It exhibits a two-branch student-labeler structure with periodic synchronization, where the labeler generates correspondences for the student. Point cloud pairs are selected at random frame interval, whose range extends with time. Labeler dirty correspondences are filtered before the speculative registration which outputs an estimated pose. Finally, correspondence rediscovery with NN-search on re-aligned input point clouds recovers clean correspondence labels.
• Figure 3: The dirty correspondence labels generated by closer-range labeler (Left: $B=1$; Right: $B=10$) on farther-apart point clouds (Left: $d=10m$; Right: $d=30m$) in KITTI Geiger2012CVPR before spatial filter. Correct ones are colored green and false ones red. Close-to-LiDAR features are less generalizable to farther pairs than far-from-LiDAR features.
• Figure 4: Visual groundings for our hypothesis on KITTI Geiger2012CVPR. (a) Density of close-to-LiDAR points are more sensitive to movement than far-from-LiDAR points. (b-d) Cosine similarity of correspondences with its distance to two LiDARs, $d_1, d_2$, under (b) $I\in [1,1]$, (c) $I\in [1,15]$, and (d) $I\in [1,30]$.
• Figure 5: Comparison between finetuning from WOD and training from scratch for EYOC, with the first 5% to 100% of unlabelled KITTI, where both RR on $d\in [40,50]$ and mRR are displayed. The horizontal axis is in log scale. Finetuning exhibits more stability before 20%, while training from scratch performs better after 50%.