README.md

---
pretty_name: "MVFGA-MoCap Dataset"
tags:
  - video
  - 3d
  - motion-capture
  - multi-view
  - face-animation
  - gesture-animation
license: other
license_name: "mvfga-mocap-research-only-license"
license_link: "https://huggingface.co/datasets/arjavanmardi/MVFGA/blob/main/LICENSE"
gated: true
extra_gated_heading: "You need to agree to share your contact information to access this dataset"
extra_gated_description: "This repository is publicly accessible, but you have to accept the conditions to access its files and content."
extra_gated_prompt: "By agreeing, you accept to share your contact information (email and username) with the repository authors and confirm that you will use the dataset only for non-commercial research in accordance with the license."
extra_gated_button_content: "Agree and access repository"
extra_gated_fields:
  "First Name": text
  "Last Name": text
  "Affiliation": text
  "Email": text
  "Phone Number": text
  "Country": country
---

# MVFGA-MoCap Dataset

<p align="center">
  <img src="teaser.png" alt="MVFGA-MoCap dataset teaser" width="100%">
</p>

<p align="center">
  <a href="https://dfki-av.github.io/MVFGA/"><strong>Project Page</strong></a>
  ·
  <a href="https://diglib.eg.org/server/api/core/bitstreams/e49bfeeb-e6f3-4391-ba4f-52e09bc18386/content"><strong>Paper</strong></a>
</p>

## Overview

**MVFGA-MoCap** is a multi-view upper-body motion-capture dataset collected for face and gesture animation research. It contains synchronized RGB videos of participants performing facial expressions, hand gestures, and natural upper-body communication motions.

The capture setup described in the accompanying work used Nikon D3200 DSLR cameras recording at 25 FPS with studio lighting. The cameras were arranged around each participant, and recordings were manually synchronized using clap gestures at the beginning and end of each sequence. Camera calibration was performed with OpenCV using checkerboard observations.

Cameras `001` through `015` capture the main frontal and side views. Cameras `016` and `017` are rear-side cameras included to provide additional viewpoints for triangulation and to improve 3D keypoint estimation, mesh fitting, and reconstruction when body parts are occluded from the frontal cameras.

## Dataset Contents

This release contains 10 subject folders. Each subject folder has the following structure:

```text
subject_xxx/
  videos/
    001.mp4
    ...
    017.mp4
  camera_params.json
```

The videos are portrait-oriented RGB MP4 files. The released files are stored at `1080 × 1920` resolution and `25 FPS`.

## Camera Parameters

Each `camera_params.json` file contains calibration data with the following entries:

- `intrinsics`: a `3 × 3` camera matrix for each calibrated camera.
- `dist`: lens-distortion coefficients.
- `world_2_cam`: a `4 × 4` world-to-camera extrinsic transformation.

The video filename corresponds to the camera ID. For example:

```text
videos/001.mp4 -> camera_params.json["intrinsics"]["001"]
videos/001.mp4 -> camera_params.json["world_2_cam"]["001"]
```

Calibration matrices are provided for cameras `001` through `015`. Videos `016.mp4` and `017.mp4` contain rear-side views, but matching calibration matrices are not included in the provided `camera_params.json` files.

## Access

Access is granted automatically after a logged-in Hugging Face user completes the access form and accepts the dataset conditions.

The submitted information is shared with the dataset repository owners for access administration and research-use tracking.

## Download

First accept the access conditions on the dataset page. Then authenticate with Hugging Face:

```bash
pip install -U huggingface_hub
hf auth login
```

Download the complete repository:

```bash
hf download arjavanmardi/MVFGA \
  --repo-type dataset \
  --local-dir MVFGA-MoCap
```

You can also download individual files from the **Files and versions** tab after access has been granted.

## Intended Use

The dataset is intended solely for non-commercial academic and scientific research related to topics such as:

- multi-view human capture;
- face and gesture animation;
- upper-body avatar reconstruction;
- 3D keypoint estimation and triangulation;
- mesh fitting and dynamic Gaussian representations.

## Restrictions

The following uses are not permitted without explicit written permission from the dataset owners:

- commercial use;
- redistribution or re-hosting of the dataset or derived copies that expose the original recordings;
- identifying or attempting to identify participants;
- profiling participants;
- impersonating participants or creating deceptive representations of them;
- using the dataset in ways that violate applicable law, institutional policy, or research-ethics requirements.

See the included [LICENSE](https://huggingface.co/datasets/arjavanmardi/MVFGA/blob/main/LICENSE) file for the complete terms.

## Citation

Please cite the accompanying paper when using this dataset:

```bibtex
@article{10.1111:cgf.70567,
  journal = {Computer Graphics Forum},
  title = {{Multi-View Face and Gesture Animation with Dynamic Gaussians}},
  author = {Javanmardi, A. and Jeetmal, V. K. and Millerdurai, C. and Pagani, A. and Stricker, D.},
  year = {2026},
  publisher = {The Eurographics Association},
  ISSN = {1467-8659},
  DOI = {10.1111/cgf.70567}
}
```

## Project Page

Further information, qualitative results, and the accompanying work are available on the [MVFGA project page](https://dfki-av.github.io/MVFGA/).

## Contact

For questions about access, licensing, or permitted uses, contact the dataset owners through the repository's **Community** tab or the contact information listed on the project page.