SynthPar aims to facilitate the development and evaluation of face recognition models, with the goal of closing the gap in performance across all demographic groups.
It provides 2 key resources:
-
A conditional StyleGAN2 generator that allows users to create synthetic face images for specified attributes like skin color and sex.
-
A dataset of ~8.9M synthetic face images with diverse skin colors and 2 sexes (Male, Female), built with VGGFace2 dataset and labels.
The dataset can be loaded from the HuggingFace repository:
from datasets import load_dataset
dataset = load_dataset("pravsels/synthpar")
To download a subset of the dataset, use the hf_dataset_dload.py script. But first, you need to setup a conda environment.
Change permissions for install_conda_env.sh and execute it:
chmod u+x ./install_conda_env.sh
./install_conda_env.sh
Activate the environment:
conda activate synthpar
Run the download script and select from subsets ST1 through ST8, which are subsets with specific skin tone regions and sex.
python hf_dataset_dload.py
Enter ST type (e.g. ST1, ST2 ... ST8):
To generate images, you need to setup a docker container.
To build and run the docker container, use docker_build.sh and run_docker_container.sh scripts respectively.
Once the container is running, the generation script can be run with the desired configuration:
python generation_script.py -c configs/ST2.yaml
This generates faces and several variations in pose and expression per face.
To generate variations in lighting on top of these, please run:
python network_demo_512.py -c configs/ST2.yaml
By default, it generates 7 lighting variations per image.
The code, dataset and model are released under the MIT license.