Virtual Garment Using Joint Landmark Prediction and Part Segmentation

摘要

We present a novel approach that constructs 3D virtual garment models from photos. Our approach only requires two images as input, one front and one back. We first apply a multi-task learning network that jointly predicts fashion landmarks and parses a garment image into semantic parts. The predicted landmarks are used for deforming a template mesh to generate 3D garment model. The semantic parts are utilized for extracting color textures for the model.