ICFFR

  This project is for the fair face recognition method IC-FFR and the NFW dataset for individual and national bias evaluation.

Train

  Part of the codes are referenced to TFace Projects. The training data can be prepared in the format of tfrecord following that project instructions. After that, you can start the training by:

  cd icffr/

  bash local_train.sh

  The training configurations can be modified according to your environment in configs/config_icffr.yaml.

Test

Data Preparation

  Download the images of our NFW from GoogleDrive, and put the whole images directory into ./data/

  cd ./data/

  unzip national_pospairs.zip

  unzip national_negpairs.zip

  cd images/

  unzip images.zip

Model Preparation

  Download the pretrained model of our IC-FFR from GoogleDrive, and put all the .pth files into ./model

Individual Bias Evaluation

  Run the script of eval_individual.py, and the result will be saved in ./results directory in a .csv file.

  Here are the individual TPR and FPR bias comparisons at different global FPRs:

 

 

National Bias Evaluation

  Run the script of eval_national.py, and the result will be saved in ./results directory in a .csv file.

  Here are the national bias comparisons by IR34 network trained on BalancedFace dataset:

 

RFW Bias Evaluation

  Here are the bias comparisons on RFW by official bias evaluation script.

 

BFW Bias Evaluation

  Here are the bias comparisons on BFW pairing all the accessible image samples in BFW dataset (Not only the official pairs).