FairFace

README

Human Face Recognition and Facial Attributes Extraction

• Reference

This repo is forked from FairFace: Face Attribute Dataset for Balanced Race, Gender, and Age for Bias Measurement and Mitigation. We modified the code slightly for actual deployment and production purpose.

• Required packages
  • pytorch1.10
  • dlib
    • Compile dlib for GPU (need to use GNC version no more than 8)

pip by default only install the cpu version of dlib, if need to deploy on cuda environment, need to install dlib from source. Please follow the below steps of installation.

Since our environment by default use GNV version 9, need to specify the gcc-8 version, otherwise will have error when building the program. Also if there is error showing the cannot find cudnn, need to specify the folder where cudnn.h locates.

# install here ~/anaconda3/pkgs/dlib
git clone https://github.com/davisking/dlib.git /anaconda3/pkgs
cd dlib
mkdir build
cd build
# test build: 
CC=gcc-8 cmake .. -DDLIB_USE_CUDA=1 -DUSE_AVX_INSTRUCTIONS=1 -DCMAKE_PREFIX_PATH=/usr/local/cuda-11.2/targets/x86_64-linux/include
cmake --build .
# build and install
python setup.py install --set USE_AVX_INSTRUCTIONS=1 --set DLIB_USE_CUDA=1 -set CUDA_HOST_COMPILER=/usr/bin/gcc-8 -set CMAKE_PREFIX_PATH=/usr/local/cuda-11.2/targets/x86_64-linux/include

• Model Architecture

  • Face recognition: In the forked repo, they used pretrain Maximum of Margin Object Detection model in dlib for face recognition.

  • Facial attribute extraction: In the forked repo, they trained a model using ResNet34 as backbone for facial attribute extraction. However, no further details about the model is disclosed in the original paper.

• How to Run the Program

1. First need to download the pretrain model here and put them under the FairFace/resnet_model folder.

2. Then run the following command in terminal, specifying:

   • --input_path as base path to store the input images
   • --input_csv as the path of Faster R-CNN prediction result of human body
   • --face_path as the path to store the detected human face images
   • --save_path as the path to store model output face_attributes.csv
   • --model_path as the path to the chosen resnet model
   • --face_size as the size of cropped human face
   • --progress as the number of images printed during model prediction progress.

python predict.py --input_path="../HKTVMall_data" --input_csv="../human_detection/result/frcnn_result.csv" --face_path="detect_faces" --save_path="result/face_attributes.csv" --model_path="res34_fair_align_multi_4_20190809.pt" --face_size=300 --progress=20

• Performance on subset of images

We first tested the model performance on a subset of 463 images.