Self-Supervised Point Set Local Descriptors for Point Cloud Registration

Descriptors play an important role in point cloud registration. The current state-of-the-art resorts to the high regression capability of deep learning. However, recent learning-based descriptors require different levels annotation and selection patches, which make model hard migrate new scenarios. In this work, we learn local registration for clouds a self-supervised manner. each iteration training, input network is merely one unlabeled cloud. Thus, whole training requires no manual patches. addition, propose involve keypoint sampling into pipeline, further improves performance our model. Our experiments demonstrate descriptor achieve even better than supervised model, while being easier train requiring data labeling.