Unpaired Translation of 3D Point Clouds with Multi-part Shape Representation

Abstract

Unpaired shape translation is an emerging task for intelligent shape modelling and editing. Recent methods for 3D shape transfer use single- or multi-scale latent codes but a single generator to generate the whole shape. The transferred shapes are prone to lose control of local details. To tackle the issue, we propose a parts-to-whole framework that employs multi-part shape representation to preserve structural details during translation. We decompose the whole shape feature into multiple part features in the latent space. These part features are then processed by individual generators respectively and transformed to point clouds. We constrain the local features of parts within the loss functions, which enable the model to generate more similar shape characteristics to the source input. Furthermore, we propose a part aggregation module that improves the performance when combining multiple point clusters to generate the final output. The experiments demonstrate that our multi-part shape representation can retain more shape characteristics compared to previous approaches.