R-PointHop: A Green, Accurate, and Unsupervised Point Cloud Registration Method

Area of Science:

• Computer Vision
• Geometric Deep Learning

Background:

• 3D point cloud registration is crucial for many applications.
• Existing methods often struggle with large rotations and require extensive training.
• Unsupervised methods offer a promising alternative to reduce data annotation burdens.

Purpose of the Study:

• To propose R-PointHop, an unsupervised method for 3D point cloud registration.
• To enhance robustness against large rotations and translations.
• To achieve efficient and accurate registration with reduced computational resources.

Main Methods:

• Determining local reference frames (LRFs) for each point.
• Extracting local-to-global hierarchical features through downsampling and neighborhood expansion.
• Establishing point correspondences using a nearest neighbor rule in feature space.
• Estimating 3D transformation from salient point correspondences.

Main Results:

• R-PointHop demonstrates effectiveness on 3DMatch, ModelNet40, and Stanford Bunny datasets.
• Achieved rotation and translation invariance in hierarchical features.
• Significantly reduced model size and training time compared to deep learning methods.
• Exhibited lower registration errors, indicating high accuracy.

Conclusions:

• R-PointHop provides a green and accurate solution for 3D point cloud registration.
• The LRF-based approach enhances robustness, especially under large rotations.
• Offers a computationally efficient alternative to existing deep learning registration techniques.