Human pose transfer via shape-aware partial flow prediction network

Human pose transfer is to transform a reference image to a new target pose. This task is challenging and existing methods still have difficulty in accurately modeling the complex spatial deformation of the body by global flow prediction. A critical reason is that it involves the shape and texture transformation of different parts, and the motion between different parts will interfere with each other. In this paper, we propose a shape-aware partial flow prediction network to implement human pose transfer. We decompose the reference pose and the target one according to the human parsing maps to predict the partial flow. The part-based flow prediction decomposes complex spatial deformation of the whole body into different parts. The shape-aware loss is to constrain the flow learning in the effective region, which provides shape information of each body part for human image synthesis. For the invisible region, we propose an image completion module, which exploits the correlation of textures to complete the features of invisible regions using that of visible regions. The synthesized results with realistic texture and accurate shape demonstrate the effectiveness of the proposed method.