GeoTransformer: Fast and Robust Point Cloud Registration With Geometric Transformer

Area of Science:

• Computer Vision
• Geometric Deep Learning
• 3D Data Processing

Background:

• Point cloud registration is crucial for 3D reconstruction and analysis.
• Keypoint-free methods offer potential, especially in low-overlap scenarios, by avoiding difficult keypoint detection.
• Existing keypoint-free methods rely on matching superpoints based on overlapping neighboring patches, requiring robust contextual geometric features.

Purpose of the Study:

• To introduce a novel method for accurate correspondence extraction in point cloud registration.
• To develop a robust feature learning approach for matching superpoints in keypoint-free registration.
• To improve the efficiency and accuracy of point cloud registration, particularly in challenging low-overlap conditions.

Main Methods:

• Proposed Geometric Transformer (GeoTransformer) for learning geometric features.
• GeoTransformer encodes pair-wise distances and triplet-wise angles for invariant and robust feature representation.
• The method matches superpoints using learned geometric features, eliminating the need for RANSAC in transformation estimation.

Main Results:

• Achieved high superpoint matching accuracy without RANSAC, resulting in a 100x acceleration.
• Demonstrated significant improvements on various benchmarks (indoor, outdoor, synthetic, multiway, non-rigid).
• Improved inlier ratio by 18–31 percentage points and registration recall by over 7 points on the 3DLoMatch benchmark.

Conclusions:

• GeoTransformer provides a highly effective and efficient solution for keypoint-free point cloud registration.
• The learned geometric features enable robust superpoint matching, even in low-overlap scenarios.
• The method significantly advances the state-of-the-art in terms of accuracy and computational speed.