Multi-Scale Semantic Transfer for Person Image Generation

Pose transfer is a task of transferring the given person pose image to the target pose image. There are mainly two problems of inaccurate generated pose and the texture dissimilarity between generated pose and target pose in the previous work. We propose the multi-scale semantic transfer network(MSTN) to generate the target pose image. It contains the semantic parsing generation module(SPGM) and multi-scale semantic transfer module(MSTM). The semantic parsing generation module(SPGM) generates target human mask with semantic information to improve structural similarity. The multi-scale semantic transfer module(MSTM) transfers pose both in high dimension and low dimension. Then it merges the information of different scales to improve texture similarity. Compared with the previous work, the posture of person and the texture such as face, clothes and hair in the images generated by our networks are more similar with the target pose image