Revert to Version 2

Training for libfacedetection in PyTorch

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

Installation

Install PyTorch >= v1.7.0 following official instruction.
Clone this repository. We will call the cloned directory as $TRAIN_ROOT.
```
git clone https://github.com/ShiqiYu/libfacedetection.train
```
Install NVIDIA DALI following official instruction: https://docs.nvidia.com/deeplearning/dali/user-guide/docs/installation.html
Install dependencies.
```
pip install -r requirements.txt
```

Note: Codes are based on Python 3+.

Preparation

Download the WIDER Face dataset and its evaluation tools.

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.

Training

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

Detection

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

Evaluation on WIDER Face

Build NMS module.

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

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

Export CPP source code

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

Export to onnx model for libfacedetection/example/opencv_dnn.

python exportonnx.py -m weights/yunet_final.pth

Compare ONNX model with other works

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:

Model	AP_easy	AP_medium	AP_hard	#Params	Params Ratio	MFlops	Froward (ms)
SCRFD0.5(ICLR2022)	0.879	0.863	0.759	631410	7.43x	184	22.3
Retinaface0.5(CVPR2020)	0.899	0.866	0.660	426608	5.02X	245	13.9
YuNet(Ours)	0.885	0.877	0.762	85006	1.0x	136	10.6

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

Citation

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

ithmz/libfacedetection.train