A clothes warping model based on reference pose map

The goal of image visual try-on is to transfer the target garment to the reference photograph and obtain a fitting result. With the popularity of deep learning in recent years, visual try-on has gradually become popular. Current technologies are often dedicated to preserving the texture of the target clothes when distorting it, and preserving as much of the human features as possible in the nontried on parts. However, in most mainstream algorithms at present, when occlusion occurs in the reference image and the posture of the human body changes, it becomes very difficult to warp clothes, fit clothes to human body and preserve the texture. In terms of this issue, we proposed some improved methods and verified them on dataset, which achieved good results. Specifically, pose map of reference human body are added as features when warping the target clothes, so that the model can fully combine the reference human body to warping the target clothes. Furthermore, the structural similarity score is added as the loss function of the clothes warping module, which can reduce the noise, keep the clothes in the same tone before and after warping as much as possible during the warping process. Compared with existing methods, our method can generate try-on images with better perceptual quality and richer details.