Towards Zero-Shot Point Cloud Registration Across Diverse Scales, Scenes, and Sensor Setups

Area of Science:

• Computer Vision
• Machine Learning
• Robotics

Background:

• Deep learning point cloud registration methods often fail in zero-shot scenarios, requiring extensive retraining or parameter tuning for new environments.
• Key limitations include fixed parameters, poor cross-domain transferability of learned detectors, and scale inconsistencies due to absolute coordinates.

Purpose of the Study:

• To develop a point cloud registration framework, BUFFER-X, capable of zero-shot generalization without domain-specific retraining or parameter tuning.
• To address limitations in scale generalization, detector transferability, and coordinate representation for robust registration across diverse datasets and sensors.

Main Methods:

• Introduced BUFFER-X, a framework utilizing a once-trained descriptor model for zero-shot generalization.
• Employed geometric bootstrapping for automatic hyperparameter estimation, distribution-aware farthest point sampling for robust feature detection, and patch-level coordinate normalization for scale consistency.
• Implemented hierarchical multi-scale matching for robust correspondence extraction across local, middle, and global receptive fields.

Main Results:

• BUFFER-X demonstrated effective zero-shot generalization across 12 diverse datasets, including object-scale, indoor, outdoor scenes, and cross-sensor LiDAR configurations.
• BUFFER-X-Lite, an optimized version, achieved a 43% reduction in computation time with comparable accuracy.
• The framework successfully generalized across varying scales, scenes, and sensor setups without parameter tuning or retraining.

Conclusions:

• BUFFER-X provides a robust and generalizable solution for point cloud registration, overcoming limitations of existing deep learning methods.
• The proposed techniques (geometric bootstrapping, distribution-aware sampling, coordinate normalization) are key to achieving effective zero-shot generalization.
• BUFFER-X and BUFFER-X-Lite offer efficient and accurate registration for diverse real-world applications.