3DPMNet: Plane Segmentation and Matching for Point Cloud Registration

A key task of indoor mobile robot navigation is to register RGB-D frames. Considering that there are many artificial planes in the indoor environment and the matching of planes is more robust than that of points, we propose a new framework to register 3D point clouds with plane correspondences in the scenes. Our framework consists of two parts: a deep neural network titled 3D Plane Matching Net (3DPMNet) which is able to segment and match planes simultaneously by taking raw point clouds as input, and a point-to-plane ICP which utilizes the plane correspondences extracted by 3DPMNet to align the two point clouds. We have generated 22000 point cloud pairs with ground-truth of plane segmentation, plane correspondences and transform matrix from ScanNet. Our experiments demonstrate that the proposed approach performs better than ICP and the recently-proposed learning-based method DCP and PRNet which utilize point correspondences. Moreover, the evaluations on plane segmentation and matching show that our 3DPMNet is able to provide accurate plane correspondences for registration.