Qin Zheng ; Yu Hao ; Wang Changjian ; Guo Yulan ; Peng Yuxing ; Ilic Slobodan ; Hu Dewen ; Xu Kai

Abstract

We study the problem of extracting accurate correspondences for point cloudregistration. Recent keypoint-free methods have shown great potential throughbypassing the detection of repeatable keypoints which is difficult to doespecially in low-overlap scenarios. They seek correspondences over downsampledsuperpoints, which are then propagated to dense points. Superpoints are matchedbased on whether their neighboring patches overlap. Such sparse and loosematching requires contextual features capturing the geometric structure of thepoint clouds. We propose Geometric Transformer, or GeoTransformer for short, tolearn geometric feature for robust superpoint matching. It encodes pair-wisedistances and triplet-wise angles, making it invariant to rigid transformationand robust in low-overlap cases. The simplistic design attains surprisinglyhigh matching accuracy such that no RANSAC is required in the estimation ofalignment transformation, leading to $100$ times acceleration. Extensiveexperiments on rich benchmarks encompassing indoor, outdoor, synthetic,multiway and non-rigid demonstrate the efficacy of GeoTransformer. Notably, ourmethod improves the inlier ratio by $18{\sim}31$ percentage points and theregistration recall by over $7$ points on the challenging 3DLoMatch benchmark.Our code and models are available at\url{https://github.com/qinzheng93/GeoTransformer}.