GPR-Net: Geometric-Positional Collaborative Point Cloud Registration Network for Repetitive Geometric Structures

Abstract

Point cloud registration is a critical task for building complete 3D models. Existing registration methods primarily focus on correspondence relationships between points, often using local geometric features to determine matches. However, this approach generates numerous false matches in scenes with repetitive geometric features due to similar structures. To address this, we propose GPR-Net, a collaborative correspondence discrimination network combining geometric and positional features of points. This method constructs a geometry branch using Dynamic Graph CNN (DGCNN) to extract multi-scale local features, while introducing a position branch to encode global spatial distributions. It employs a collaborative discrimination mechanism to select point pairs exhibiting both geometric similarity and spatial consistency. The main contributions of this work are threefold: (1) we design a dual-branch framework that integrates geometric and positional features for robust correspondence discrimination; (2) we introduce contrastive constraints in both branches to enhance feature distinctiveness and reduce mismatches in repetitive structures; (3) we demonstrate superior registration results compared to state-of-the-art descriptors through extensive experiments on ModelNet40 and noisy datasets. Experiments show our method effectively counteracts interference from similar structures, outperforming mainstream feature descriptors in registration accuracy and robustness within repetitive geometric feature scenes.