Multi-stage Network with Garments Extraction Module to Improve Image-based Virtual Try-on

In the existing image-based virtual try-on works, the garment fabric is often wrongly retained in the neck, chest and some other positions. To deal with that, a multi-stage network with garments extraction module is proposed to enhance the quality of the generated images at the aforementioned area. The proposed network consists of four modules: the garment extraction module, garment warping module, semantic layout generation module and content fusion module. Our main work are the import of the garment extraction module and the improvement of the semantic layout generation module. The proposed method is evaluated qualitatively on the virtual try-on dataset VITON and a clothing dataset collected from the Internet which is mainly consists of hoodies and low necklines. Compared with the current virtual try-on networks CP-VTON and ACGPN, the proposed approach obtains more natural images especially in some specific areas such as neck and chest. Despite the addition of extra modules, our work cost only 0.02ms extra compared to ACGPN.