It is the training program for libfacedetection. The source code is based on FaceBoxes.PyTorch and ssd.pytorch.

Visualization of our network architecture: [netron].

• Installation
• Preparation
• Training
• Detection
• Evaluation on WIDER Face
• Export CPP source code
• Export to ONNX model
• Design your own model
• Citation

1. Install PyTorch >= v1.7.0 following official instruction.

2. Clone this repository. We will call the cloned directory as $TRAIN_ROOT.

   git clone https://github.com/ShiqiYu/libfacedetection.train

3. Install NVIDIA DALI following official instruction: https://docs.nvidia.com/deeplearning/dali/user-guide/docs/installation.html

4. Install dependencies.

   pip install -r requirements.txt

Note: Codes are based on Python 3+.

1. Download the WIDER Face dataset and its evaluation tools.
2. Extract zip files under $TRAIN_ROOT/data/widerface as follows:

   $ tree data/widerface
   data/widerface
   ├── wider_face_split
   ├── WIDER_test
   ├── WIDER_train
   ├── WIDER_val
   └── trainset.json           

NOTE:
We relabled the WIDER Face train set using RetinaFace. New labels are in $TRAIN_ROOT/data/widerface/trainset.json, which is the COCO_format annotations file used in DALI dataloader.

python train.py -c ./config/yufacedet.yaml -t demo 

python detect.py -c ./config/yufacedet.yaml -m weights/yunet_final.pth --target filename.jpg 

1. Build NMS module.

   cd tools/widerface_eval
   python setup.py build_ext --inplace

2. Perform evaluation. To reproduce the following performance, run on the default settings. Run python test.py --help for more options.

   python test.py -m weights/yunet_final.pth -c ./config/yufacedet.yaml

NOTE: We now use the Python version of eval_tools from here.

Performance on WIDER Face (Val): confidence_threshold=0.3, nms_threshold=0.45, in origin size:

AP_easy=0.882, AP_medium=0.871, AP_hard=0.767

The following bash code can export a CPP file for project libfacedetection

python exportcpp.py -c ./config/yufacedet.yaml -m weights/yunet_final.pth

Export to onnx model for libfacedetection/example/opencv_dnn.

python exportonnx.py -m weights/yunet_final.pth

Inference on exported ONNX models using ONNXRuntime:

python tools/compare_inference.py ./onnx/yunet_final_dynamic_simplify.onnx --mode AUTO --eval --score_thresh 0.3 --nms_thresh 0.45

Some similar approaches(e.g. SCRFD, Yolo5face, retinaface) to inference are also supported.

With Intel i7-12700K and input_size = origin size, score_thresh = 0.3, nms_thresh = 0.45, some results are list as follow:

The compared ONNX model is avaliable in https://share.weiyun.com/nEsVgJ2v Password：gydjjs

The loss used in training is EIoU, a novel extended IoU. More details can be found in:

@article{eiou,
 author={Peng, Hanyang and Yu, Shiqi},
 journal={IEEE Transactions on Image Processing}, 
 title={A Systematic IoU-Related Method: Beyond Simplified Regression for Better Localization}, 
 year={2021},
 volume={30},
 pages={5032-5044},
 doi={10.1109/TIP.2021.3077144}
 }

The paper can be open accessed at https://ieeexplore.ieee.org/document/9429909.

We also published a paper on face detection to evaluate different methods.

@article{facedetect-yu,
 author={Yuantao Feng and Shiqi Yu and Hanyang Peng and Yan-ran Li and Jianguo Zhang}
 title={Detect Faces Efficiently: A Survey and Evaluations},
 journal={IEEE Transactions on Biometrics, Behavior, and Identity Science},
 year={2021}
 }

The paper can be open accessed at https://ieeexplore.ieee.org/document/9580485