Semantic Segmentation for 3D Human Models

Abstract

We propose a technique for semantic segmentation of 3D human models. Existing techniques for general 3D objects solely rely on geometric information, which may suffer from distinguishing, e.g., skin and thin clothes. Our method thus exploits texture information as well. Specifically, our method first generates 2D data via projection from multiple viewpoints and feeds the data to neural networks to obtain likelihoods of semantic labels, which are then back-projected onto the human model. Note that likelihoods for some polygons are not available due to low-resolution 2D data and occlusion. We finally assign semantic labels for the entire mesh via graphcut.