Garment mesh reconstruction and animation

Abstract

With great potential in online garment shopping and game animation, recent years research in garment virtual try-on becomes more and more popular. Garment virtual try-on can be divided into three parts, namely 3D garment mesh reconstruction, animation and texture transfer. Most of current research only focus on one of the sub-fields and are difficult to be combined. There is still no endto-end pipeline. In this report, we propose a network to reconstruct the 3D garment mesh that is easily to be combined with animation model and make contribution to the completeness of the end-to-end pipeline.Our network can properly predict 3D garment meshes according to various garment styles. Unlike some of the existing approaches that require complex inputs, our model works with 2D garment images or masks that are easily accessible. Unsupervised learning is implemented during training, making the data collection much easier as no labelling is required. We have successfully combined our model with existing 3D animation model. Now the network can directly work from 2D images instead of 3D garment information.