Boosting RGB-D Point Cloud Registration Via Explicit Position-Aware Geometric Embedding

The goal of point cloud registration is to determine the rigid transformation between two unaligned point clouds, in which extracting reliable correspondences from overlapping regions is the key to success. Many recent methods have attempted to boost point cloud registration by using RGB-D data. Their pivotal design is to employ a visual branch and a geometric branch to extract visual and geometric features respectively, and then fuse them for correspondence prediction. However, most of them focus on the joint learning and fusion strategies of two distinctive modality features while ignoring the impact of ambiguities caused by non-overlapping regions. Numerous feature points sampled from non-overlapping regions bring correspondence ambiguities, which degrades the performance significantly, especially in low-overlap scenarios. To this end, this paper proposes a simple yet effective performance-boosting strategy for RGB-D point cloud registration, which alleviates the interference of non-overlapping regions by explicitly embedding the overlapping cues extracted from 2D visual correspondences. Specifically, inspired by a recently proposed position-aware geometric embedding paradigm for point cloud registration, we leverage the 2D visual correspondences extracted from the visual branch to guide the geometric branch to sample feature points and to learn the local geometric dependencies among them. Furthermore, an explicit position-aware geometric embedding module is proposed to enhance the 3D spatial awareness of the geometric branch. Extensive experiments on ScanNet and 3DMatch/3DLoMatch datasets demonstrate that our model outperforms the state-of-the-art methods. The source code is publicly available at https://github.com/R-Levi/RGBD-PCR.