Coarse-to-Fine Volumetric Prediction for Single-Image 3D Human Pose Supplementary Material

In this supplementary, we provide material that could not be included in the main manuscript due to space constraints. First, Section 1 provides additional details for the exact representation of the volumetric space and the way we can obtain metric pose estimates from the voxelized estimates. Section 2 presents full results on Human3.6M using the reconstruction error for evaluation. Section 3 further examines the decoupled architecture in the case the groundtruth 2D joints are provided as input, while Section 4 focuses on quantitative evaluation on the MPII Human Pose dataset using our volumetric representation. Finally, Section 5 presents additional quantitative results and provides insights about the failure cases or our approach.