HR-Net: Point cloud registration with hierarchical coarse-to-fine regression network

We introduce a hierarchical coarse-to-fine registration network named HR-Net for point cloud registration. Our method incorporates the features of different layers and predicts the rigid transformation without using RANSAC in a coarse-to-fine strategy. Specifically, we combine the reliable features of sparse keypoints in the lowest downsampled layer and dense precise position information in the upsampling layer. The coarse transformation is obtained by using the sparse keypoints and features of the coarse registration layer. A fine registration correspondence layer is presented to generate accurate point correspondences by incorporating coarse registration features, our fine registration uses local spatial attention to aggregate the features and points instead of global attention requiring numerous computing resources. Furthermore, we have carried out extensive experiments on 3DMatch and ModelNet benchmarks. Our approach surpasses the most advanced method by a margin of 2.8 percentage points on registration recall and obtains lower registration error on the challenging 3DLoMatch benchmark.