This is the code accompanying the paper Does Data Repair Lead to Fair Models? Curating Contextually Fair Data To Reduce Model Bias published in IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2022.

Clone the repo:

git clone https://github.com/sumanyumuku98/contextual-bias.git

Create Conda env:

conda env create -f fair.yml

Activate the conda env:

conda activate fair.yml

Download the COCO 2017 annotations and create a symlink:

cd data
ln -s /path to annotations/ .

Generate a barplot for the protected attribute/class w.r.t other co-occuring classes:

python genPlt.py --data_mode train --bias_class man

The above example generates co-occuring distribution for man. The generated barplots will be saved in out_figures.

To curate contextually fair data for a protected attribute and fixed budget w.r.t to either topK, bottomK and custom co-occuring categories use:

python fairSelection.py --data_mode train --bias_class cup --budget 500 --k 10

The above example curates fair data for protected class cup with a budget of 500 and co-occuring classes being the top 10. The selected ids are saved in out_files.

If you find the work useful, do cite:

@InProceedings{Agarwal_2022_WACV,
    author    = {Agarwal, Sharat and Muku, Sumanyu and Anand, Saket and Arora, Chetan},
    title     = {Does Data Repair Lead to Fair Models? Curating Contextually Fair Data To Reduce Model Bias},
    booktitle = {Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)},
    month     = {January},
    year      = {2022},
    pages     = {3298-3307}
}

If there are any questions or concerns feel free to send a message at [email protected]