Human Pose Transfer Based on Deep Learning Models: A Survey

Human pose transfer has received considerable of academic attention because of its numerous uses in movie production, computer-graphics-based manipulations or data augmentation. Deep learning advancements have made it possible to simply generate or change visual appearances (face, pose, outfit) of a human image on demand. In this survey, we first describe the extent of human pose transfer and then examine current developments and technology trends methodically, dividing it into three major paths: Methods that use only 2D pose (joints), methods that use surface-based pose (Human Parsing) and methods that use 3D (Mesh or DensePose). 3D pose methods achieved the best score but on the cost of time while surface-based methods produce better and sharper images than the 2D methods because they provide a better pose representation. We believe that this survey sheds some insight into the prospects of Human posture transfer and serves as a useful baseline.