Human3.6M¶
- Modality: Optical motion capture (3D joint positions) + high-resolution RGB + depth + ground-truth camera calibration
- Primary Tasks: 3D human pose estimation, pose forecasting, mocap to image alignment
- Scale: 3.6 million 3D frames, 11 professional actors, 15 activity scenarios, 4 synchronized cameras
- License: Research-only; requires signed license agreement with the Human3.6M consortium
- Access: http://vision.imar.ro/human3.6m/description.php
Summary¶
Human3.6M is a cornerstone benchmark for monocular and multi-view 3D pose estimation. Captured with Vicon mocap and synchronized video, it provides high-quality ground-truth joint trajectories aligned to RGB frames, enabling supervised training and precise evaluation of 3D human reconstruction methods.
Reference Paper¶
- Catalin Ionescu et al. "Human3.6M: Large Scale Datasets and Predictive Methods for 3D Human Sensing in Natural Environments." IEEE TPAMI, 2014.
PDF
Benchmarks & Baselines¶
- Martinez et al. (Simple Baseline) - MPJPE 67.5 mm (procrustes-aligned) on Protocol #2.
- PoseFormer - MPJPE 44.3 mm on Protocol #1; Zheng et al., ICCV 2021.
- Evaluation protocols commonly reported: Protocol #1 (17 joints, MPJPE), Protocol #2 (Procrustes-aligned MPJPE), Protocol #3 (prediction on unnormalized coordinates).
Tooling & Ecosystem¶
- Official toolkit with MATLAB evaluation scripts available upon request.
- Human3.6M Toolkit for data preprocessing and baseline models.
- MMPose maintains configs and evaluation utilities.
Known Challenges¶
- Data access approval may take several days; download links can be slow.
- Mixed lighting conditions and motion blur challenge monocular models.
- Protocol confusion is common-explicitly document which evaluation protocol you follow.
Cite¶
@article{ionescu2014human3,
title = {Human3.6M: Large Scale Datasets and Predictive Methods for 3D Human Sensing in Natural Environments},
author = {Ionescu, Catalin and Papava, Dragos and Olaru, Vlad and Sminchisescu, Cristian},
journal = {IEEE Transactions on Pattern Analysis and Machine Intelligence},
year = {2014}
}