A 3D Mesh-Based Lifting-and-Projection Network for Human Pose Transfer

Human pose transfer has typically been modeled as a 2D image-to-image translation problem. This formulation ignores the human body shape prior in 3D space and inevitably causes implausible artifacts, especially when facing occlusion. To address this issue, we propose lifting-and-projection framework to perform mesh space. The core of our is foreground generation module, that consists two novel networks: lifting-and-projection network (LPNet) an appearance detail compensating (ADCNet). leverage prior, LPNet exploits topological information learn expressive visual representation for target person preserve texture details, ADCNet further introduced enhance feature produced by with source image. Such design module enables model better handle difficult cases such those occlusions. Experiments on iPER Fashion datasets empirically demonstrate proposed effective outperforms existing image-to-image-based mesh-based methods task both self-transfer cross-transfer settings.