Unsupervised 3D Learning for Shape Analysis via Multiresolution Instance Discrimination

We propose an unsupervised method for learning a generic and efficient shape encoding network different analysis tasks. Our key idea is to jointly encode learn point features from unlabeled 3D clouds. For this purpose, we adapt HRNet octree-based convolutional neural networks with fused multiresolution subnetworks design simple-yet-efficient Multiresolution Instance Discrimination (MID) loss the features. takes cloud as input output both After training, concatenated simple task-specific back-ends fine-tuned evaluate efficacy generality of our set tasks, including classification, semantic segmentation, well registration With back-ends, demonstrates best performance among all methods achieves competitive supervised methods. fine-grained segmentation on PartNet dataset, even surpasses existing by large margin.