Recover Human Body Structure from Still Image



This is an on-going project by Leon Gu and Professor Takeo Kanade. Our goal is to recover the structure of human body, a nineteen-points skeleton from still images automatically. Dealing with large deformation of articulated structures, making use of noisy image evidence and separating structure variations from pose changes are three key challenges. We address these challenges from a Bayesian perspective. Our focus is on how to incorporate a multi-modal shape prior learned from CMU Mocap database with various levels of image cues learned from labeled human body images.