GeoTransformer

Why we included this project

For 3D perception and LiDAR pipelines, GeoTransformer offers a clean, well-documented reference for point cloud registration. Instead of detecting repeatable keypoints, its superpoint matching network encodes pair-wise distances and triplet-wise angles, keeping it reliable in low-overlap scans and invariant to rigid transforms. The matching accuracy is reported high enough to estimate the alignment transform without RANSAC, which the authors say speeds the pipeline up about a hundred times. Pretrained models for 3DMatch, 3DLoMatch, KITTI, and ModelNet40 are included, so you can reproduce the paper's results or fine-tune for your own sensor setup. If you're working on SLAM, mapping, or scan-to-model alignment and want a modern baseline, this repository is a concrete starting point.