A novel partial-to-partial registration method based on sampling network

Point cloud registration is mainly to estimate a rigid transformation between point clouds. The traditional optimization-based registration method requires a good initial position, and it is easy to fall into a local optimal solution. Some learning-based methods are introduced to reduce the dependence on the initial transformation, but they cannot handle partial-to-partial registration tasks. This paper proposes a learning-based registration method for partial-to-partial scenario. The local geometry is encoded into the feature representation of each point. A transformer network is used to enhance attention features. A designed sampling network down-sample key matching points and their corresponding features. The rigid transformation is calculated according to virtual correspondence by a singular value decomposition layer. The ModelNet40 dataset and Stanford 3D Scanning models are used to test the registration performance. Experimental results show that the proposed method achieves better registration accuracy than traditional methods, and it is robust to any initial transformation and noise.