Unsupervised 3D Learning for Shape Analysis via Multiresolution Instance Discrimination

Although unsupervised feature learning has demonstrated its advantages toreducing the workload of data labeling and network design in many fields,existing unsupervised 3D learning methods still cannot offer a generic networkfor various shape analysis tasks with competitive performance to supervisedmethods. In this paper, we propose an unsupervised method for learning ageneric and efficient shape encoding network for different shape analysistasks. The key idea of our method is to jointly encode and learn shape andpoint features from unlabeled 3D point clouds. For this purpose, we adaptHR-Net to octree-based convolutional neural networks for jointly encoding shapeand point features with fused multiresolution subnetworks and design asimple-yet-efficient Multiresolution Instance Discrimination (MID) loss forjointly learning the shape and point features. Our network takes a 3D pointcloud as input and output both shape and point features. After training, thenetwork is concatenated with simple task-specific back-end layers andfine-tuned for different shape analysis tasks. We evaluate the efficacy andgenerality of our method and validate our network and loss design with a set ofshape analysis tasks, including shape classification, semantic shapesegmentation, as well as shape registration tasks. With simple back-ends, ournetwork demonstrates the best performance among all unsupervised methods andachieves competitive performance to supervised methods, especially in taskswith a small labeled dataset. For fine-grained shape segmentation, our methodeven surpasses existing supervised methods by a large margin.