Transcoding across 3D Shape Representations for Unsupervised Learning of 3D Shape Feature

ABSTRACT Unsupervised learning of 3D shape feature is a challenging yet important problem for organizing a large collection of 3D shape models that do not have annotations. Recently proposed neural network-based approaches attempt to learn meaningful 3D shape feature by autoencoding a single 3D shape representation such as voxel, 3D point set, or multiview 2D images. However, using single shape representation isn't sufficient in training an effective 3D shape feature extractor, as none of existing shape representation can fully describe geometry of 3D shapes by itself. In this paper, we propose to use transcoding across multiple 3D shape representations as the unsupervised method to obtain expressive 3D shape feature. A neural network called Shape Auto-Transcoder (SAT) learns to extract 3D shape features via cross-prediction of multiple heterogeneous 3D shape representations. Architecture and training objective of SAT are carefully designed to obtain effective feature embedding. Experimental evaluation using 3D model retrieval and 3D model classification scenarios demonstrates high accuracy as well as compactness of the proposed 3D shape feature. The code of SAT is available at https://github.com/takahikof/ShapeAutoTranscoder