• Introduction
  • Functions
  • Test_environment
• Evaluation
  • Widerface
  • Parameter_and_flop
  • Speed
• How_to_use
  • Installation
  • Training
  • Evaluation
  • Convertor
  • C++_inference
• References

This project provides a serias of lightweight face detectors with landmarks which can be deployed on mobile devices.

• Modify the anchor size of Retinaface-mobile0.25
• Add landmarks estimation to Face-Detector-1MB

• Train/test/evaluation/ncnn/tensorflow/tflite/C++ inference of Retinaface-mobile0.25
• Train/test/evaluation/ncnn/tensorflow/tflite/C++ inference of Face-Detector-1MB slim and RFB version
• Add 5 landmarks estimation to Face-Detector-1MB
• Support the inference using pytorch/ncnn/tensorflow/tflite

• Ubuntu18.04
• Python3.7
• Pytorch1.2
• CUDA10.0 + CUDNN7.5

• Evaluation result on wider face val (input image size: 320*240）

• Evaluation result on wider face val (input image size: 640*480)

• Test speed on RK3399 using tflite format

Input image size: 320*240

Input image size: 640*480

1. git clone this project

2. Pytorch version 1.1.0+ and torchvision 0.3.0+ are needed.

3. Codes are based on Python 3

1. The dataset directory as follows:

  ./data/widerface/
    train/
      images/
      label.txt
    val/
      images/
      wider_val.txt

ps: wider_val.txt only include val file names but not label information.

2. We provide the organized dataset we used as in the above directory structure.

Link: from google cloud or baidu cloud Password: ruck

1. Before training, you can check network configuration (e.g. batch_size, min_sizes and steps etc..) in data/config.py and train.py.

 adjuct lmk_weight and loc_weight to improve the landmarks accuracy

2. Train the model using WIDER FACE:

CUDA_VISIBLE_DEVICES=0 python train.py --network mobile0.25 or 
CUDA_VISIBLE_DEVICES=0 python train.py --network slim or
CUDA_VISIBLE_DEVICES=0 python train.py --network RFB

If you don't want to train, we also provide a trained model on ./weights

mobilenet0.25_Final.pth 
RBF_Final.pth
slim_Final.pth

1. Generate txt file

python test_widerface.py --trained_model weight_file --network mobile0.25 or slim or RFB

2. Evaluate txt results. Demo come from Here

cd ./widerface_evaluate
python setup.py build_ext --inplace
python evaluation.py

3. You can also use widerface official Matlab evaluate demo in Here

• Convert pytorch to onnx/ncnn/caffe/tensorflow/tflite
• Converting script on convertor folder

1. Generate onnx file

python convert_to_onnx.py --trained_model weight_file --network mobile0.25 or slim or RFB

2. Onnx file change to ncnn(*.param and *.param)

cp *.onnx ./Detector_cpp/Face_Detector_ncnn/tools
cd ./Detector_cpp/Face_Detector_ncnn/tools
./onnx2ncnn face.param face.bin

3. Simplify onnx file

pip install onnx-simplifier
python-m onnxsim input_onnx_model output_onnx_model

4. Convert to Caffe

python convertCaffe.py

5. Convert to Tensorflow

python demoCaffe.py
python froze_graph_from_ckpt.py

6. Convert to Tensorflow lite

python convert_to_tflite.py

• C++ inference code for ncnn/tf/tflite on Detector_cpp folder

1. Build Project(set opencv path in CmakeList.txt)

mkdir build
cd build
cmake ..
make -j4

2. run

./FaceDetector *.jpg
./FaceDetectorFolder [folder path]

• FaceBoxes
• Retinaface (mxnet)
• Retinaface (pytorch)
• Ultra-Light-Fast-Generic-Face-Detector-1MB
• Face-Detector-1MB-with-landmark

@inproceedings{deng2019retinaface,
title={RetinaFace: Single-stage Dense Face Localisation in the Wild},
author={Deng, Jiankang and Guo, Jia and Yuxiang, Zhou and Jinke Yu and Irene Kotsia and Zafeiriou, Stefanos},
booktitle={arxiv},
year={2019}