Seonglae ChoExisting methods require manual adjustment of voxel size and search radius for each new environment, learned keypoint detectors easily fail when the domain changes, and using raw coordinates leads to performance drops when scale differences increase.
- Adaptive geometric bootstrapping determines optimal voxel size and search radius automatically for each scene using PCA-based 'sphericity' analysis.
- Uses Farthest Point Sampling (FPS) instead of unstable deep keypoints.
- Normalizes patch coordinates to [-1,1] range to eliminate scale dependency, and extracts and matches features at local, middle, and global scales.
- Selects final inliers by checking cross-scale consistency among transformations estimated from correspondences at each scale.
Achieves high success rates in zero-shot without prior tuning across 11 dataset benchmarks covering RGB-D, LiDAR, indoor, outdoor environments, and various sensors, countries, and scales.