Abstract
Point cloud registration is a fundamental task in robotics and 3D vision, yet it remains challenging under conditions of low overlap and arbitrary initial rotations. We present SE3ET, a novel framework that leverages SE(3)-equivariant transformers to achieve robust alignment. By embedding geometric symmetry directly into the neural architecture, SE3ET achieves state-of-the-art performance and demonstrates exceptional generalization across diverse environments without the need for extensive data augmentation.
Introduction
Point cloud registration involves determining the relative transformation between two 3D scans, a critical component for autonomous navigation and mapping. Traditional deep learning approaches often struggle with two primary real-world hurdles: low overlap between scans (often as low as 10-30%) and sensitivity to initial orientations. While many models rely on rotation-invariant features, these often discard vital geometric information. SE3ET addresses these limitations by utilizing equivariant features that preserve geometric structure while ensuring consistent responses across all orientations.
Methodology
The SE3ET framework is built upon the principle of SE(3)-equivariance, where internal features rotate predictably in response to input rotations.
Equivariant Backbone
The model employs an Efficient SE(3)-Equivariant Point Network (E2PN) as its backbone. Unlike previous methods that use complex icosahedral groups, E2PN discretizes the rotation group into an octahedral group (24 rotations). This approach is computationally efficient and integrates seamlessly with voxel-based downsampling techniques.
Equivariant Transformer Design
Standard transformers typically ignore the geometric structure of 3D tensors. SE3ET introduces four specialized transformer designs—ESA, ICA, ACA, and RCA—that allow the attention mechanism to reason about spatial locations and orientations simultaneously. This ensures that the transformer layers maintain the equivariant property throughout the feature extraction process.
Coarse-to-Fine Alignment
The pipeline follows a hierarchical approach: it first identifies correspondences between "superpoints" (coarse patches) to establish a global alignment and then refines the transformation using dense point-to-point correspondences.
Experimental Evaluation
SE3ET was evaluated on several rigorous benchmarks:
- 3DMatch & 3DLoMatch: Indoor RGB-D datasets, with 3DLoMatch specifically testing low-overlap (10-30%) scenarios.
- Rotated Benchmarks: Harder variants of the indoor datasets where test pairs are subjected to arbitrary 3D rotations.
- KITTI: A large-scale outdoor LiDAR dataset. A custom low-overlap version was created by selecting frames separated by at least 10 meters.
Results and Discussion
SE3ET demonstrates significant improvements in robustness and precision compared to existing state-of-the-art models.
Performance Metrics
On the Rotated 3DLoMatch benchmark, SE3ET-I achieved a Registration Recall of 70.8%, surpassing GeoTransformer (68.5%) and the equivariant model YOHO (65.9%). In outdoor settings (KITTI), SE3ET achieved a Relative Translation Error of just 4.7 cm and a Relative Rotation Error of 0.21°, the lowest among all compared methods.
Generalization and Efficiency
One of the most striking results is the model's zero-shot generalization capability. When trained on indoor data (3DMatch) and tested on outdoor data (KITTI), SE3ET maintained a 75.9% recall. In contrast, the non-equivariant RoITr model's performance dropped to 20.9%. Furthermore, the SE3ET-I2 variant remains computationally efficient, processing pairs in approximately 105ms with fewer parameters than many non-equivariant competitors.
Conclusion
By incorporating symmetry as a structural prior, SE3ET provides a robust solution for point cloud registration in challenging environments. The model's ability to handle arbitrary rotations and low overlap without exhaustive data augmentation marks a significant step forward for reliable robotic perception. Future work will focus on reducing the dependency on RANSAC for final alignment and integrating the framework into full SLAM systems.
Summary updated