Consensus-Based Optimization for 3D Human Pose Estimation in Camera Coordinates

3D human pose estimation is frequently seen as the task of estimating poses relative to root body joint. Alternatively, we propose a method in camera coordinates, which allows effective combination 2D annotated data and straightforward multi-view generalization. To that end, cast problem view frustum space estimation, where absolute depth prediction joint estimations are disentangled. Final predictions obtained coordinates by inverse projection. Based on this, also present consensus-based optimization algorithm for from uncalibrated images, requires single monocular training procedure. Although our indirectly tied intrinsics, it still converges cameras with different intrinsic parameters, resulting coherent up scale factor. Our improves state art well known datasets, reducing error 32% most common benchmark. We reported results position error, achieving 80 mm 51 multi-view, average. Source code available at https://github.com/dluvizon/3d-pose-consensus .