Semantic Segmentation and Labeling of 3D Garments

As large collections of 3D garments continue to grow, analyzing and exploring shape variations is significant but challenging. In this paper, we propose a semi-supervised learning method for semantic segmentation and labeling of 3D garments. The key idea in this work is to address the data challenge for 3D garment analysis using semi-supervised learning method which can label parts in various 3D garments. We first develop an objective function based on Conditional Random Field (CRF) model to learn the prior knowledge of garment components from a set of training examples. Then, we segment 3D garments into five component prototypes related to top, bottom, sleeve, accessory and one-piece, respectively. And we modify the Joint Boost to automatically cluster the segmented components without requiring manual parameter tuning. The purpose of our method is to relieve the manual segmentation and labeling of components in 3D garment collections. The experimental results demonstrate our method is effective and comparable to human work.