Ego-Exo4D Human Meshes Dataset

Multi-view reconstruction of 3D human body and hand pose on Ego-Exo4D — 523 hours of video across 8 activities.

Browse the data Get the dataset Install & run

Dataset overview

—Takes processed

—Total video hours (5 cams)

—Per-camera hours

—Activity types

Interactive viewer

Documentation

Installation

Environment setup, model checkpoints, and a verified test run.

Downloading the data

Raw Ego-Exo4D dataset, plus the pre-computed release on Hugging Face.

Running the pipeline

Single takes, batch jobs, resuming, and the low-level optimizer.

Data format

The released npz file structure.

Rendering meshes

Turn a results npz into rendered SMPL mesh videos.

Code released under the MIT License. Builds on SLAHMR (Ye et al., CVPR 2023).

@inproceedings{ye2023slahmr,
    title={Decoupling Human and Camera Motion from Videos in the Wild},
    author={Ye, Vickie and Pavlakos, Georgios and Malik, Jitendra and Kanazawa, Angjoo},
    booktitle={IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
    month={June},
    year={2023}
}