LEARNING POINT CLOUD REGISTRATION

Abstract

Point cloud registration refers to the task of finding the spatial transformation that aligns two or more point clouds. It is an important problem with many applications in robotics and virtual or augmented reality. In this thesis comprising four works, we ­­­propose several ways to learn from data to improve point cloud registration performance. We first propose RPM-Net, which learns to refine the alignment between two partially overlapping point clouds in an iterative manner using learned hybrid features. The second work, 3DFeat-Net, addresses the problem of learning local features for correspondence-based registration approaches in a weakly supervised manner, through the use of approximate GPS/INS poses. Next, we propose REGTR, which is a RANSAC-free correspondence-based registration method.­­ Finally, we tackle the problem of registration of more than two point clouds by designing a learned view-graph optimization approach using recurrent graph neural networks.