Self-Supervised Few-Shot Learning on Point Clouds

The increased availability of massive point clouds coupled with their utilityin a wide variety of applications such as robotics, shape synthesis, andself-driving cars has attracted increased attention from both industry andacademia. Recently, deep neural networks operating on labeled point clouds haveshown promising results on supervised learning tasks like classification andsegmentation. However, supervised learning leads to the cumbersome task ofannotating the point clouds. To combat this problem, we propose two novelself-supervised pre-training tasks that encode a hierarchical partitioning ofthe point clouds using a cover-tree, where point cloud subsets lie within ballsof varying radii at each level of the cover-tree. Furthermore, ourself-supervised learning network is restricted to pre-train on the support set(comprising of scarce training examples) used to train the downstream networkin a few-shot learning (FSL) setting. Finally, the fully-trainedself-supervised network's point embeddings are input to the downstream task'snetwork. We present a comprehensive empirical evaluation of our method on bothdownstream classification and segmentation tasks and show that supervisedmethods pre-trained with our self-supervised learning method significantlyimprove the accuracy of state-of-the-art methods. Additionally, our method alsooutperforms previous unsupervised methods in downstream classification tasks.