Deep learning-based interest point detector for 3D point clouds

Abstract

Interest point (keypoint) detection for 3D point clouds is the problem of finding stable points that are well repeatable in the 3D point cloud under arbitrary rigid transformations. These detected keypoints play essential roles in many autonomous driving and robotics applications such as 3D point cloud-based odometry, place recognition, or 3D point cloud-based localization. In these applications, the detected keypoints in different frames are further used to extract representative features for matching, computing transformations, and estimating locations. Although there are many conventional interest point detection methods for 3D point clouds, they usually need hand-crafting for specific datasets and cannot be generalized easily to other datasets. While 2D (image) keypoint detectors have been quite successful, this is not so for 3D keypoint detectors. In this project, we investigate the challenges faced by 3D keypoint detectors and develop deep learning approaches to detect the keypoints.