Person image generation via regional style rectification

Person image generation is a challenging task and has been used in many people centered applications. Current person image generation methods mainly focus on transferring the texture features of the source image to the target pose, but ignore the constraint effect of the source image on the generated results. In this paper, we propose a regional style rectification network for person image generation, which aims to use the style information of source images to improve the textures of the generated person images. Within the proposed model, we design a regional style compensation module to rectify the shape-independent style information of the generated image to the source by producing a regional style residual map. Furthermore, we propose a regional style consistency loss to strengthen the style rectification, which realizes the direct supervision of source images over target image generation, ignoring the shape differences between the source and target images. Besides, we use a dual-attention-based texture transformation module which exploits the correlations between source and target images to better preserve the texture features of source images. The experiment results for pose transfer and attribute editing demonstrate the effectiveness of the proposed model.