A hybrid algorithm with inlier-guided Hough voting for point cloud registration

Point cloud registration is a fundamental yet challenging task due to the contamination by large numbers of outliers in correspondences. Following the classical model generation-and-selection solution, many methods are dedicated to exploring effective outlier removal strategies to obtain inlier correspondences and generate candidate hypotheses. However, these methods usually suffer from inaccurate model selection, resulting in the failure of point cloud registration. In this paper, we propose a hybrid point cloud registration framework that considers both the quality of hypothesis generation and the accuracy of model selection. For model generation, a global compatibility-based inlier sampling is performed to obtain outlier-free consensus and produce reliable hypotheses. Further, we present an inlier-guided Hough voting network for model selection, which incorporates the inlier count prior into the voting process and conducts maximum likelihood estimation to select the best transformation accurately. Extensive experiments on 3DMatch, 3DLoMatch, and KITTI datasets indicate that the proposed point cloud registration method achieves state-of-the-art performance.