End-to-End Learning the Partial Permutation Matrix for Robust 3D Point Cloud Registration

Even though considerable progress has been made in deep learning-based 3D point cloud processing, how to obtain accurate correspondences for robust registration remains a major challenge because existing hard assignment methods cannot deal with outliers naturally. Alternatively, the soft matching-based have proposed learn matching probability rather than assignment. However, this paper, we prove that these an inherent ambiguity causing many deceptive correspondences. To address above challenges, propose partial permutation matrix, which does not assign corresponding points outliers, and implements prevent ambiguity. proposal poses two new problems, i.e. algorithms can only solve full rank matrix desired is defined discrete space, non-differentiable. In response, design dedicated soft-to-hard (S2H) procedure within pipeline consisting of steps: solving (S-step) projecting (H-step). Specifically, augment profit before augmented cropped achieve final matrix. Moreover, guarantee end-to-end learning, supervise learned but propagate gradient instead. Our S2H be easily integrated frameworks, verified representative frameworks including DCP, RPMNet, DGR. Extensive experiments validated our method, creates state-of-the-art performance.