LOCO-PCR: A Novel Outlier Removal Method with Local Correction for Feature-based Point Cloud Registration

Finding reliable correspondences between two point clouds is crucial for feature-based point cloud registration (PCR). In this paper, we present a novel outlier removal method, which locally corrects (LOCO) the initial correspondences to generate more correct correspondences, denoted as LOCO-PCR. First, we use spectral matching technique to find some potential inliers, called super seeds. For each super seed, we first find its k neighbors and then construct new local correspondences by local feature matching. Second, we propose a local hypothesis generation module to generate more accurate hypotheses. We construct global and local consensus sets based on the spatial compatibility of initial global correspondences and new local correspondences, respectively. Subsequently, we input each global and local consensus set into the weighted SVD algorithm to generate candidate rigid transformations. Finally, we propose a cross hypothesis selection strategy to select the best model from the global and local hypotheses as the final result. Our method can be combined with different local 3D features without training, further improving the accuracy and stability of feature-based point cloud registration.