Mutual A Priori based Completion for Low Overlapping Point Clouds Registration

This work presents a new registration method that especially designed for low-overlapping partial point clouds. Based on the assumption that the partial point clouds to be registered belong to the same target, the proposed Mutual a Priori based Completion (MPC) method uses the partial point clouds to be registered as the a priori for each other to expend each individual point clouds. The main challenge of this mutual reference approach is that the partial clouds without spatial alignment cannot provide reliable completion reference for each. Based on the mutual information maximization, a progressive completion structure is developed to make the alignment in the spatial, feature and completion space respectively. Experiments on public datasets show encouraging results. Especially for the ultra-low overlap cases, compared with the state-of-the-art (SOTA) models, the size of overlapping regions can be increased by about 14.0%, and the rotation and translation error can be reduced by 30.8% and 57.7% respectively.