Learning to Analyze and Synthesize Human-Centric Images

The ability to automatically analyze and synthesize human-centric data has been garnering significant interest in recent years. It was propelled by the increasing demands of real-world applications. Traditional computer-generated image technology requires massive amounts of human work and is hard to apply to casually captured daily photos. Recent deep learning-based methods enable data-driven approaches. My research targets real-world data and investigates data-driven methods to infer human body structure and generate realistic novel images. In particular, my research covers three parts: 1) human dense pose estimation, learning to map a 2D image onto a 3d body surface; 2) human pose transfer, learning to repose a human from a single 2D image; 3) human-centric image synthesis, learning to hallucinate both a human body and its surroundings.