TASFNet: Robust Point Cloud Correspondence via Topological and Semantic Integration

In the field of point cloud registration, the extraction of correspondences of low-overlap point clouds is still a difficult challenge, and its results have a decisive influence on the quality of the registration. Although current methods have made significant progress in solving this problem, the abnormal matching problem is still a major bottleneck, limiting the accuracy and reliability of registration. Based on this, we innovatively proposed TASF (Topological and Semantic Framework), which captures the topological structure and semantic features of point clouds at the same time, thereby revealing the deep information hidden in the point cloud, and can achieve high-quality point correspondence even under low overlap conditions. The superiority of TASF has been fully verified on two point cloud registration datasets, 3DMatch and 3DLoMatch. Especially in the test of low-overlap datasets such as 3DLoMatch, our method significantly improved the matching ratio of points in the point cloud, up to 8.8 percentage points higher than the existing methods. This result verifies the feasibility of our method.