Face Detection Development

This repository is dedicated to face detection development. It supports training, evaluation, and ONNX model conversion. The first implementation included in this repo is SCRFD.

Environment setup

MacOS

• Pyenv

• Set up xz for pyenv to install Python 3.10.16

  brew install xz
  echo 'export PYTHON_CONFIGURE_OPTS="--with-lzma=$(brew --prefix xz)' > ~/.zshrc
  exec $SHELL
  pyenv install 3.10.16
  pyenv virtualenv 3.10.16 face_detection_dev
  pyenv activate face_detection_dev

• Python packages

  pip install wheel
  pip install torch==2.4.1 torchvision # must fix for mps bug, see https://github.com/pytorch/pytorch/issues/142344
  MAX_JOBS=4 MMCV_WITH_OPS=1 pip install git+https://github.com/open-mmlab/[email protected]
  pip install -r docker/requirements_macos.txt

Linux

• Pyenv

  pyenv install 3.10.16
  pyenv virtualenv 3.10.16 face_detection_dev
  pyenv activate face_detection_dev

• Python packages

  pip install wheel
  pip install torch==2.5.1 torchvision
  MAX_JOBS=4 MMCV_WITH_OPS=1 pip install git+https://github.com/open-mmlab/[email protected]
  pip install -r docker/requirements_linux.txt

Docker

• Build Docker image

  cd docker
  bash build.sh
  cd -

• Run Docker container

  bash docker/run_docker.sh $CMD

Train

prepare dataset

We use gdown to download the dataset. Please install it first.

pip install gdown

Then download the dataset and untar it.

gdown 1nr5QhnDIeiMApzK_fjYsh0BjD3_OwPgw
tar cvf face_detection.tar

without docker

python -u train.py ${config}

with docker

bash docker/run_docker.sh python train.py ${config}

To onnx

ipython -- torch2onnx.py --config ${config} --ckpt_fpath ${ckpt_fpath} --onnx_fpath ${onnx_fpath}

Benchmarks

Widerface Evaluation

ipython -- benchmark/widerface.py --onnx_fpath ${onnx_fpath}

NIST Evaluation

ipython -- benchmark/nist.py --onnx_fpath ${onnx_fpath}