Unsupervised non-rigid point cloud registration based on point-wise displacement learning

Registration of deformable objects is a fundamental prerequisite for many modern virtual reality and computer vision applications. However, due to the difficulties of acquiring labeled datasets and the inherent irregular deformation, non-rigid registration for 3D scanner-captured data remains challenging. This paper proposes an unsupervised non-rigid 3D point cloud registration network based on the self-attention mechanism. Specifically, considering the registration as the result of point drifts between the source and target shapes, a Transformer-based encoder-decoder module is utilized to estimate the point displacements. Additionally, a symmetric registration procedure is adopted with regularization loss to manage the regular deformation of points, ultimately producing reasonable registration results for real-world deformable objects. Experiments are conducted on public and synthesized datasets which simulate diversiform non-rigid 2D or 3D deformations. Numerical and qualitative experimental results demonstrate that the proposed network achieves outstanding performance and is robust in scenes with multiple interferences.