kentoo-zbcast/libfacedetection.train

The training program for libfacedetection for face detection and 5-landmark detection.

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].

Contents

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

Installation

1. Install PyTorch >= v1.0.0 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 dependencies.

pip install -r requirements.txt

Note: Codes are based on Python 3+.

Training

1. Download WIDER FACE dataset, place the images under this directory:

$TRAIN_ROOT/data/WIDER_FACE_rect/images

and create a symbol link to this directory from

$TRAIN_ROOT/data/WIDER_FACE_landmark/images

2. Train the model using WIDER FACE:

cd $TRAIN_ROOT/tasks/task1/
python3 train.py

Detection

cd $TRAIN_ROOT/tasks/task1/
./detect.py -m weights/yunet_final.pth --image_file=filename.jpg

Evaluation on WIDER Face

1. Enter the directory.

cd $TRAIN_ROOT/tasks/task1/

2. Create a symbolic link to WIDER Face. $WIDERFACE is the path to WIDER Face dataset, which contains wider_face_split/, WIDER_val, etc.

ln -s $WIDERFACE widerface

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

mkdir results
python test.py

4. Download and run the official evaluation tools. NOTE: Matlab required!

# download
wget http://shuoyang1213.me/WIDERFACE/support/eval_script/eval_tools.zip
# extract
unzip eval_tools.zip
# run the offical evaluation script
cd eval_tools
vim wider_eval.m # modify line 10 and line 21 according to your case
matlab -nodesktop -nosplash -r "run wider_eval.m;quit;"

Performance on WIDER Face (Val)

Run on default settings: scales=[1.], confidence_threshold=0.3:

AP_easy=0.852, AP_medium=0.823, AP_hard=0.646

Export CPP source code

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

cd $TRAIN_ROOT/tasks/task1/
./exportcpp.py -m weights/yunet_final.pth -o output.cpp

Export to onnx model

Export to onnx model for libfacedetection/example/opencv_dnn.

cd $TRAIN_ROOT/tasks/task1/
python exportonnx.py -m weights/yunet_final.pth

Design your own model

You can copy $TRAIN_ROOT/tasks/task1/ to $TRAIN_ROOT/tasks/task2/ or other similar directory, and then modify the model defined in file: tasks/task2/yufacedetectnet.py .

Citation

Our paper, which introduces a novel loss named Extended IoU (EIoU), is coming out soon. We trained our model using the EIoU loss and obtained a performance boost, see Performance on WIDER Face (Val) for details. Stay tune for the release of our paper!