Virtual Try-on via Matching Relation with Landmark

Virtual try-on based on image synthesis aims to combine the customer's photo with in-shop clothes to acquire a try-on image. The key to generating a realistic try-on image is whether the in-shop clothes is spatially aligned with the customer's body. Prior methods usually directly adopt the spatial transformation network to complete the clothing deformation, but they cannot generate high-quality try-on images when facing the customer's posture changes or complex clothes pattern. To address it, we propose a virtual try-on network based on landmark constraint (LCVTON) in this work. Specifically, we notice the corresponding relationship between the clothes feature points and the customer body feature points, making the clothes match more closely with the customer's body. The matching of the feature points enables us to introduce the landmark constraint into the spatial transformer network for naturally and smoothly warping clothes. Moreover, we construct a refinement network and introduce the landmark constraint into it to preserve the texture details of clothes. We conducted experiments on the try-on dataset and compared our method with existing methods. Both qualitative and quantitative results demonstrate the superiority of our method compared to existing state-of-the-art method.