Learning-based Point Cloud Registration: A Short Review and Evaluation

Weixuan TANG, Danping Zou, Ping Li

Abstract: Point cloud registration is an important task for range scan alignment, pose estimation, and localization. Traditional point cloud registration methods rely on hand-craft descriptors, which are sometimes not so descriptive and make the pose solver easy to fail because of false matchings. Recently, many researchers seek to improve the traditional method by deep learning-based approach. In this paper, we summarize the main pipeline of point cloud registration in traditional and learning-based approaches. Then we review some of the recent start-of-art methods, mainly in the end-to-end learning approach. We also review the criteria used to evaluate the registration performance and give complete testing results, some of which are not provided by those papers.