Unsupervised cycle-consistent person pose transfer

Abstract

Person pose transfer, i.e., transferring the pose of a given person to a target pose, is a challenging task due to the complex interplay of appearance, pose, and background. Most of the previous works adopted the supervised framework and required paired person images with the same identity and different poses, which largely limits their applications. Besides, the background of the generated image may be altered from the original one due to some over-fitting issues, which is unfavorable for the pose transfer task. To tackle the above problems, we propose an unsupervised cycle-consistent person pose transfer approach. It is trained with unpaired cross-identity person images and can well preserve the background information. Compared with previous methods, our proposed approach achieves better results in the cross-identity person pose transfer task and similar results in self-identity one. Moreover, our method can serve as an effective data augmentation scheme for person recognition tasks, which is validated by extensive experiments on pedestrian re-identification and detection.