PARE-Net: Position-Aware Rotation-Equivariant Networks for Robust Point Cloud Registration

TL;DR

A contrastive rotation loss is presented to enhance the robustness of rotation-equivariant features against data degradation, and a position-aware convolution is proposed, which can better learn spatial information of local structures.

Abstract

Learning rotation-invariant distinctive features is a fundamental requirement for point cloud registration. Existing methods often use rotation-sensitive networks to extract features, while employing rotation augmentation to learn an approximate invariant mapping rudely. This makes networks fragile to rotations, overweight, and hinders the distinctiveness of features. To tackle these problems, we propose a novel position-aware rotation-equivariant network, for efficient, light-weighted, and robust registration. The network can provide a strong model inductive bias to learn rotation-equivariant/invariant features, thus addressing the aforementioned limitations. To further improve the distinctiveness of descriptors, we propose a position-aware convolution, which can better learn spatial information of local structures. Moreover, we also propose a feature-based hypothesis proposer. It leverages rotation-equivariant features that encode fine-grained structure orientations to generate reliable model hypotheses. Each correspondence can generate a hypothesis, thus it is more efficient than classic estimators that require multiple reliable correspondences. Accordingly, a contrastive rotation loss is presented to enhance the robustness of rotation-equivariant features against data degradation. Extensive experiments on indoor and outdoor datasets demonstrate that our method significantly outperforms the SOTA methods in terms of registration recall while being lightweight and keeping a fast speed. Moreover, experiments on rotated datasets demonstrate its robustness against rotation variations. Code is available at https://github.com/yaorz97/PARENet.

Figures (12)

• Figure 1: Experimental results on 3DMatch and 3DLoMatch. Our method significantly outperforms state-of-the-art methods w.r.t. registration recall (RR) while maintaining fast speed and lightweight.
• Figure 2: Comparison of our feature-based hypothesis proposer, RANSAC and LGR on 3DLoMatch.
• Figure 3: The framework of our method. Given point clouds $\mathcal{P}$ and $\mathcal{Q}$, a hierarchical backebone based on PARE-Conv is used to extract patch-level and point-level features. Then, we follow a coarse-to-fine scheme to obtain point correspondences using rotation-invariant features. Finally, the feature-based hypothesis proposer leverages the matched rotation-equivariant features ${\tilde{ \mathbf{F} }}^{\mathcal{P}}_{x_{jn}}$ and ${\tilde{ \mathbf{F}} }^{\mathcal{Q}}_{y_{jn}}$, which encode fine-grained structure orientations, to generate multiple reliable hypotheses. The best hypothesis is selected and refined as the final solution ($\mathbf{R}^*$, $\mathbf{t}^*$).
• Figure 4: An illustration of incorrect alignment caused by the ambiguity of using coordinates to align. This problem can be solved by using rotation-equivariant features that contain fine-grained local orientation information.
• Figure 5: Qualitative results. Our method can successfully align low-overlapped pairs with higher IR and PIR (Patch IR). Moreover, GeoTrans qin2022geometric and PEAL peal incorrectly align point clouds with opposite orientations (the first and second rows), our method successfully addresses this issue because the equivariant features contain orientation information of local structure, alleviating this ambiguity.