SingleReg: An Unsupervised Registration Method for Point Cloud With Good Generalization Performance

Point cloud registration is a pivotal procedure in the processing of point cloud data, where matching of corresponding points is essential for high-precision alignment. To reduce the likelihood of incorrect matches, this study introduces an unsupervised registration network that operates effectively with a minimal set of corresponding points. Unlike prior methods that depend on similarity measures of feature descriptors, our approach leverages the geometric consistency within the point cloud itself to establish correspondence, significantly enhancing the network’s generalizability. Comprehensive experimental evaluations on the ModelNet40, Stanford, and 3DMatch datasets reveal that our methodology surpasses existing state-of-the-art unsupervised approaches. In addition, it demonstrates robust generalization capabilities, maintaining high performance in cross-dataset evaluations without the necessity for retraining.