Pose transforming network: Learning to disentangle human posture in variational auto-encoded latent space

Abstract

This paper proposes a novel deep conditional generative model for human pose transforms. To generate the desired pose-transformed images from a single image, a variational inference model is formulated to disentangle human posture semantics from image identity (human personality, background etc.) in variational auto-encoded latent space. A deep learning architecture is then proposed to realize the formulated variational inference model. In addition, a new loss function for the proposed training method is designed to enable pose information and identity information to be separated completely in the latent space. The proposed model is validated experimentally by demonstrating its pose-transforming capability, outperforming the existing conditional generative model.