GMNet: Low overlap point cloud registration based on graph matching

Abstract

Point cloud registration quality relies heavily on accurate point-to-point correspondences. Although significant progress has been made in this area by most methods, low-overlap point clouds pose challenges as dense point topological structures are often neglected. To address this, we propose the graph matching network (GMNet), which constructs graph features based on the dense point features obtained from the first point cloud sampling and the superpoints’ features encoded with geometry. By using intra-graph and cross-graph convolutions in local patches, GMNet extracts deeper global information for robust correspondences. The GMNet network significantly improves the inlier ratio for low-overlap point cloud registration, demonstrating high accuracy and robustness. Experimental results on public datasets for objects, indoor, and outdoor scenes validate the effectiveness of GMNet. Furthermore, on the low-overlap 3DLoMatch dataset, our registration recall rate remains stable at 72.6%, with the inlier ratio improving by up to 9.9%.