Point Cloud Registration via Reconstruction with Local Geometry Information Aggregation

Abstract

Point cloud registration is a fundamental yet challenging task in computer vision and robotics. While framing it as a reconstruction problem has shown promise, traditional reconstruction approaches rely on positional encodings to encode positional information, which inadequately capture the intricate geometric and positional relationships between point cloud pairs, leading to suboptimal registration results. To overcome this limitation, we introduce a Local Geometry Information Aggregation (LGIA) module that effectively captures both global positional context and fine-grained local geometric details. Furthermore, to better utilize the geometric and positional information inherent in point clouds, we propose a salient point sampling strategy that increases the proportion during the downsampling process. Subsequently, we assign these salient points higher weights during patch matching. Experiments demonstrate the superiority of our method on 3DMatch, 3DLoMatch and KITTI datasets, which achieves SOTA results at three metrics of FRM, RRT and RTE, and especially 98.5% FRM on the 3DMatch dataset.