Posern: A 2D Pose Refinement Network For Bias-Free Multi-View 3D Human Pose Estimation

We propose a new 2D pose refinement network that learnsto predict the human bias in the estimated 2D pose. There arebiases in 2D pose estimations that are due to differences be-tween annotations of 2D joint locations based on annotatorsƒ??perception and those defined by motion capture (MoCap) sys-tems. These biases are crafted into publicly available 2D posedatasets and cannot be removed with existing error reductionapproaches. Our proposed pose refinement network allowsus to efficiently remove the human bias in the estimated 2Dposes and achieve highly accurate multi-view 3D human poseestimation.