Unsupervised 3D Shape Representation Learning using Normalizing Flow

Xiang Li, Congcong Wen, Hao Huang; Proceedings of the Asian Conference on Computer Vision (ACCV), 2022, pp. 1124-1141

Learning robust and compact shape representation learning plays an important role in many 3D vision tasks. Existing supervised learning-based methods have achieved remarkable performance, meanwhile requiring large-scale human-annotated datasets for model training. Self-supervised/unsupervised methods provide an attractive solution to this issue that can learn shape representations without the need for ground truth labels. In this paper, we introduce a novel self-supervised method for shape representation learning using normalizing flows. Specifically, we build a model upon a variational normalizing flow framework where a sequence of normalizing flow layers are adopted to model exact posterior latent distribution and enhance the representation power of the learned latent code. To further encourage inter-shape separability and intra-shape compactness among a batch of shapes, we design a contrastive-center loss that performs metric learning on features on a hypersphere. We validate the representation learning ability of our model on downstream classification tasks. Experiments on ModelNet40/10, ScanobjectNN, and ScanNet datasets demonstrate the superior performance of our method compared with current state-of-the-art methods.