Self-Supervised Deep Learning on Point Clouds by Reconstructing Space

Point clouds provide a flexible and natural representation usable incountless applications such as robotics or self-driving cars. Recently, deepneural networks operating on raw point cloud data have shown promising resultson supervised learning tasks such as object classification and semanticsegmentation. While massive point cloud datasets can be captured using modernscanning technology, manually labelling such large 3D point clouds forsupervised learning tasks is a cumbersome process. This necessitates methodsthat can learn from unlabelled data to significantly reduce the number ofannotated samples needed in supervised learning. We propose a self-supervisedlearning task for deep learning on raw point cloud data in which a neuralnetwork is trained to reconstruct point clouds whose parts have been randomlyrearranged. While solving this task, representations that capture semanticproperties of the point cloud are learned. Our method is agnostic of networkarchitecture and outperforms current unsupervised learning approaches indownstream object classification tasks. We show experimentally, thatpre-training with our method before supervised training improves theperformance of state-of-the-art models and significantly improves sampleefficiency.