Distribution-motivated 3D Style Characterization Based on Latent Feature Decomposition

Xinwei Huang, Shuai Li, Shoulong Zhang, Aimin Hao, Hong Qin

Abstract: The style characterization for 3D shapes remains mathematically elusive in spite of rapid research progress in 3D vision. Previous unorthodox approaches frequently identify the 3D shape style with geometric attributes (e.g., size, texture, sub-parts, skeleton) and their complex combinations, which hinder the generalizability of style analysis and synthesis tasks. The central idea of our current research is hinged upon the style and content/structure decoupling based on the latent feature decomposition. In this paper, the 3D shape style is implicitly defined as certain channels of the latent features learned by a shape reconstruction auto-encoder. Based on this definition, we devise a novel Style Feature Decomposition Module (SFDM) to automatically disentangle style from content and structure in the latent space. In particular, the SFDM adopts the Earth Mover’s Distance (EMD), characterizing the style-related channels with large inter-class distribution differences between the source and target shapes. Meanwhile, in order to preserve the source content information, we keep the most stable feature channels based on the intra-class distribution stability. The SFDM is implemented in a feed-forward strategy without any assistance from the correspondence or segmentation sub-tasks. Comprehensive experiments have confirmed that, our newly-proposed SFDM can successfully decompose the shape style representation from the latent space, and it can naturally enable and facilitate various 3D down-stream applications such as style-driven shape generation, deformation, and interpolation.