Towards Multi-pose Guided Virtual Try-on Network

Contribution

• We introduce a novel task of virtual try-on conditioned on multiple poses, and collect a new dataset that covers different poses and various clothes.
• We propose a novel Multi-pose Guided Virtual Try-On Network (MG-VTON) that handles large pose varia- tions by disentangling the warping of clothes appear- ance and the pose manipulation in multiple stages. Specifically, we propose a pose-clothes guided human parsing network to first synthesize the human parsing with the desired clothes and pose, which effectively guides the virtual try-on to achieve reasonable results via the correct region parts.
• We design a Warp-GAN that integrates human pars- ing with geometric matching to alleviate blurry issues caused by the misalignment among different poses.
• A pose-guided refinement network is further proposed to adaptively controls the composition mask according to different poses, which learns to recover details and remove artifacts.