IPAB Workshop-20/08/2020

Title:Learning Global 3D Human Motion from 2D Keypoint Sequences without paired Supervision

 

Abstract:The canonical task of 3d human pose estimation from images or videos is mainly concerned with root-relative poses. A subject's joint can freely move relative to the hip of the body but the body itself does not move. In a well defined environment like one with a fixed camera and a known ground plane, we can exploit assumptions about global motion to jointly learn global motion and local poses from 2D keypoints. We show that this is possible with a dataset of unrelated 3D motion sequences as the only 3D supervision using adversarial learning.