AMASS¶

Modality: Unified MoCap parameter sequences (SMPL/SMPL+H/DMPL parameters)
Primary Tasks: Motion synthesis, pose estimation, imitation learning, generative modeling
Scale: ~16,000 minutes, 344 subjects, 40+ MoCap datasets merged
License: AMASS custom research license (non-commercial, attribution required)
Access: https://amass.is.tue.mpg.de/

Summary¶

AMASS (Archive of Motion Capture as Surface Shapes) consolidates dozens of motion-capture datasets into a common SMPL parameterization, providing temporally consistent body poses and shapes. It serves as a foundational corpus for human motion synthesis and as pretraining data for pose estimation models that operate in parametric body-space.

Reference Paper¶

Armin Mahmood et al. "AMASS: Archive of Motion Capture as Surface Shapes." ICCV, 2019. PDF

Benchmarks & Baselines¶

VPoser prior trained on AMASS underpins many downstream pose/shape models; Mahmood et al., ICCV 2019.
Motion VAE based on AMASS achieves high-fidelity sequence generation; Kanazawa et al., CVPR 2019 (HMR follow-ups).
No single official leaderboard; evaluations often use training/validation splits by source dataset to prevent leakage.

Tooling & Ecosystem¶

Official AMASS processing code for downloading and converting sequences.
SMPLify-X leverages AMASS for pose/shape priors.
MoCap Act provides dataset loaders aligned with AMASS segments.

Known Challenges¶

Licensing varies per constituent dataset; ensure you honor original dataset terms when redistributing derivatives.
Some sequences lack foot-contact labels; contact detection requires post-processing.
Requires large storage (~43 GB compressed) and relies on SMPL model assets (need separate license from MPI).

Cite¶

@inproceedings{mahmood2019amass,
  title     = {AMASS: Archive of Motion Capture as Surface Shapes},
  author    = {Mahmood, Armin and Ghorbani, Navid and Pons-Moll, Gerard and others},
  booktitle = {Proceedings of the IEEE International Conference on Computer Vision},
  year      = {2019}
}