Hosts the Multiface dataset, which is a multi-view dataset of multiple identities performing a sequence of facial expressions.
Hosts the Multiface dataset, which is a multi-view dataset of multiple identities performing a sequence of facial expressions.
Our dataset consists of high quality recordings of the faces of 13 identities, each captured in a multi-view capture stage performing various facial expressions. An average of 12,200 (v1 scripts) to 23,000 (v2 scripts) frames per subject with capture rate at 30 fps. Each frame includes roughly 40 (v1) to 160 (v2) different camera views under uniform illumination, yielding a total dataset size of 65TB. We provide the raw captured images from each camera view at a resolution of 2048 × 1334 pixels, tracked meshes including headposes, unwrapped textures at 1024 × 1024 pixels, metadata including intrinsic and extrinsic camera calibrations, and audio. This repository hosts the code of downloading the dataset and building a Codec Avatar using a deep appearance model. To learn more about how the dataset is captured and how different model architectures can influence performance, you may refer to our Technical Report.
To download our data, first clone this repository and install dependencies
git clone https://github.com/facebookresearch/multiface
cd multiface
pip3 install -r requirements.txtSince the full dataset takes terabytes of storage, one may wish to download partially. If you want to view the example assets, you may download the mini-dataset (16.2 GB)
python3 download_dataset.py --dest "/path/to/mini_dataset/" --download_config "./mini_download_config.json"The download_config argument points to the configuration file specifying assets to be downloaded, options include:
| Variable | Type | Default |
|---|---|---|
| entity | list of string | All the entity will be downloaded |
| image | boolean | Raw images of entities selected will be downloaded |
| mesh | boolean | Tracked mesh of entities selected will be downloaded |
| texture | boolean | Unwrapped texture of entities selected will be downloaded |
| metadata | boolean | Metadata of entities selected will be downloaded |
| audio | boolean | Audio of entities selected will be downloaded. The first available frame is aligned with the start of the audio file, however, missing frames in images need to be handled for alignment |
| expression | list of string | All the facial expression (contains both v1 and v2 scripts) will be downloaded |
The configuration to download all assets can be found at download_config.json.
To run training and render the 3D faces, please refer to our full installation guide.
To learn more on selecting model architecture, camera split and expression split for training and testing set, please refer to quick start.
Thanks to all the people who has helped generate and maintain this dataset!
If you use any data from this dataset or any code released in this repository, please cite the technical report (https://arxiv.org/abs/2207.11243)
…Multiface is CC-BY-NC 4.0 licensed, as found in the LICENSE file.
No open issues yet, or sync has not completed.