nghorbani/amass

What it solves

AMASS provides a unified framework and parameterization for multiple optical marker-based motion capture (mocap) datasets. It solves the problem of fragmented human motion data by consolidating various sources into a common format, making it easier to use for animation, visualization, and training deep learning models.

How it works

The project uses the MoSh++ pipeline to fit the SMPL+H body model to mocap data. It represents human motion as surface shapes, supporting an extended shape space (16 betas and 8 DMPLs). The repository provides tools and Jupyter notebooks to animate SMPL+H, SMPL, and SMPL-X body models using the AMASS data.

Who it’s for

Researchers and developers working in computer vision, animation, and deep learning who need high-quality, standardized human motion capture data for training or visualization.

Highlights

• Unified database of human motion from diverse mocap datasets.
• Support for multiple body models including SMPL+H, SMPL, and SMPL-X.
• Includes pre-defined non-overlapping train/validation/test data splits.
• Tools for producing synthetic mocap using DFaust registrations.