Learning 3D Point Cloud Registration as a Single Optimization Problem

Rintaro Yanagi, Atsushi Hashimoto, Naoya Chiba, Shusaku Sone, Jiaxin Ma, Yoshitaka Ushiku; Proceedings of the Asian Conference on Computer Vision (ACCV), 2024, pp. 3292-3309

We present a simple yet effective technique to boost the performance of 3D point cloud registration. It only requires replacing the conventional hand-crafted differentiable matching module with an uncertainty-aware one. Conventional methods pass a distance (or similarity) matrix to the matching module as a deterministic input, ignoring any uncertainty in upstream distance calculation. In other words, existing works always consider the optimality of the feature extractor and matching module separately, which is sub-optimal. We propose to use a trainable matching network as the uncertainty-aware matching module and connect it with a feature extractor in a non-deterministic way. The matching network is trained with the feature extractor as a single probabilistic process in this way, yielding a joint-optimal solution. Experimental results have demonstrated that our strategy significantly boosts the performance of recent SOTA methods under versatile conditions, including rigid/non-rigid and whole/partial point cloud registration datasets.