Multi-features guidance network for partial-to-partial point cloud registration

The recent extraction of hybrid features improves point cloud registration performance by emphasizing more integrated information. However, hybrid features ignore the large dimensional differences, big semantic gaps, and mutual interference between the shape features and spatial coordinates. This paper proposes a novel Multi-Features Guidance Network (MFGNet) for partial-to-partial point cloud registration to overcome the intrinsic flaws of hybrid features, which leverages the shape features and the spatial coordinates to account for correspondences searching independently. The proposed network mainly includes four parts: keypoints’ feature extraction, correspondences search, correspondences credibility computation, and singular value decomposition (SVD), among which correspondences search and correspondences credibility computation are the cores of the network. Specifically, the correspondences search module utilizes the shape features and the spatial coordinates to guide correspondences matching independently and fusing the matching results to obtain the final matching matrix. Moreover, based on the conflicted relationship between the two matching matrices, the correspondences credibility computation module scores each correspondence pair’s reliability, which can reduce the impact of mismatched or non-matched points significantly. Empirical experiments on the ModelNet40 dataset validate the effectiveness of the proposed MFGNet, which achieves 0.19 $$^\circ$$ , 0.24 $$^\circ$$ and 1.3 $$^\circ$$ mean absolute errors for rotation matrix and 0.0010, 0.0011, and 0.0068 mean absolute errors for translation vectors, respectively, under the settings of unseen point clouds, unseen categories, and Gaussian noise.