GTINet: Global Topology-Aware Interactions for Unsupervised Point Cloud Registration

Abstract

Point cloud registration is a critical task in various 3D applications. Supervised approaches are restricted by the difficulty and cost of acquiring ground-truth annotations. Thus, unsupervised point cloud registration has emerged as a promising alternative. However, existing unsupervised methods often overlook the importance of feature interactions, leading to feature matching ambiguity. To address these challenges, we propose an unsupervised point cloud registration framework termed Global Topology-aware Interactions Network (GTINet), which contains a global structural relations (GSR) module and a contextual topological interactions (CTI) module. The GSR module transforms local features into global features through global graph convolutions. Based on the obtained global features, the CTI module learns geometric feature similarities and relative positional knowledge for both the source and target point clouds. The CTI module further learns contextual feature interactions through topology-aware attention layers. By improving the discriminativeness of features, our GTINet reduces the feature matching ambiguity caused by local structural similarity. Extensive experiments demonstrate that our method achieves state-of-the-art unsupervised registration performance on the ModelNet40, 7Scene, and KITTI datasets. Our work provides a novel perspective for conducting unsupervised point cloud registration. We will release our code for future research.