Self-Supervised Learning of Local Features in 3D Point Clouds

Abstract

We present a self-supervised task on point clouds, in order to learn meaningful point-wise features that encode local structure around each point. Our self-supervised network, operates directly on unstructured/unordered point clouds. Using a multi-layer RNN, our architecture predicts the next point in a point sequence created by a popular and fast Space Filling Curve, the Morton-order curve. The final RNN state (coined Morton feature) is versatile and can be used in generic 3D tasks on point clouds. Our experiments show how our self-supervised task results in features that are useful for 3D segmentation tasks, and generalize well between datasets. We show how Morton features can be used to significantly improve performance (+3% for 2 popular algorithms) in semantic segmentation of point clouds on the challenging and large-scale S3DIS dataset. We also show how our self-supervised network pretrained on S3DIS transfers well to another large-scale dataset, vKITTI, leading to 11% improvement. Our code is publicly available.1