Boundary-aided Human Body Shape and Pose Estimation from a Single Image for Garment Design and Manufacture

Abstract

Current virtual clothing design applications mainly use predefined virtual avatars which are created by professionals. The models are unrealistic as they lack the personalised body shapes and the simulation of human body muscle and soft tissue. To address this problem, we firstly fit the state-of-the-art parametric 3D human body model, SMPL, to 2D joints and boundary of the human body which are detected using CNN methods automatically. Considering the scenario of virtual dressing where people are usually in stable poses, we define a stable pose prior from CMU motion capture (mocap) dataset for further improving accuracy of pose estimation. Accurate estimation of human body shape and poses provides manufacturers and designers with more comprehensive human body measurements, which put a step forwards clothing design and manufacture through Internet.