A Representation Separation Perspective to Correspondence-Free Unsupervised 3-D Point Cloud Registration

3D point cloud registration in remote sensing field has been greatly advanced by deep learning based methods, where the rigid transformation is either directly regressed from two clouds (correspondences-free approaches) or computed learned correspondences (correspondences-based approaches). Existing correspondences-free methods generally learn holistic representation of entire cloud, which fragile for partial and noisy clouds. In this paper, we propose a unsupervised (UPCR) method separation perspective. First, model input as combination pose-invariant pose-related representation. Second, used to relative pose wrt "latent canonical shape" source target respectively. Third, obtained above poses. Our not only filters out disturbance but also robust partial-to-partial noise. Experiments on benchmark datasets demonstrate that our achieves comparable if better performance than state-of-the-art supervised methods.