Human Upper Body 3D Pose Estimation

Abstract

Solving 3D human pose estimation challenge from a monocular video lacks large, diverse datasets, particularly videos with just upper-body parts visible. As a step towards mitigating this issue, we retrained a 3D pose estimation model on upper-body joints and tested its performance on an augmented dataset created from a standard video database. The performance of the re-trained model was shown to be better than the full-body pose model which was trained to handle occlusions.