GeoTransformer
GeoTransformer is a PyTorch implementation of a deep learning model for fast and robust point cloud registration, originally presented at CVPR 2022. It addresses the challenge of finding accurate correspondences in low-overlap scenarios by utilizing a keypoint-free approach. Instead of detecting repeatable keypoints, the method operates on downsampled superpoints derived from neighboring patches. The core innovation is the Geometric Transformer architecture, which learns geometric features by encoding pairwise distances and triplet-wise angles. This design ensures invariance to rigid transformations and robustness where traditional methods fail. The simplistic yet effective architecture achieves such high matching accuracy that it eliminates the need for RANSAC during alignment transformation estimation, resulting in a computational speedup of approximately 100 times. The tool significantly improves performance on standard benchmarks like 3DLoMatch, increasing inlier ratios by 17 to 30 percent and registratio