Blazeface

实现功能

• Blazeface的训练/测试/评估/ncnn C++推理
• Face-Detector-1MB slim和RFB版本的训练/测试/评估/ncnn C++推理
• 人脸5个关键点检测
• 支持onnx导出
• 网络parameter和flop计算

网页版demo

• run in web browser! 👉 https://github.com/zineos/ncnn-webassembly-blazeface

带有关键点检测的超轻量级人脸检测器（小于1MB）

提供了一系列适合移动端部署包含关键的人脸检测器: 在Face-Detector-1MB基础上实现了Blazeface 并添加了关键点检测和ncnn C++部署功能.

测试的运行环境

• Ubuntu20.04
• Python3.8.5
• Pytorch1.7
• CUDA11.2 + CUDNN8

精度

Widerface测试

• 在wider face val精度（单尺度输入分辨率：320*240）

方法	Easy	Medium	Hard
libfacedetection v1（caffe）	0.65	0.5	0.233
libfacedetection v2（caffe）	0.714	0.585	0.306
version-slim(原版)	0.765	0.662	0.385
version-RFB(原版)	0.784	0.688	0.418
version-slim(our)	0.795	0.683	0.34.5
version-RFB(our)	0.814	0.710	0.363
Retinaface-Mobilenet-0.25(our)	0.811	0.697	0.376
Blazeface	0.782	0.715	0.394

• 在wider face val精度（单尺度输入分辨率：640*480）

方法	Easy	Medium	Hard
libfacedetection v1（caffe）	0.741	0.683	0.421
libfacedetection v2（caffe）	0.773	0.718	0.485
version-slim(原版)	0.757	0.721	0.511
version-RFB(原版)	0.851	0.81	0.541
version-slim(our)	0.850	0.808	0.595
version-RFB(our)	0.865	0.828	0.622
Retinaface-Mobilenet-0.25(our)	0.873	0.836	0.638
Blazeface	0.877	0.816	0.639

ps: 测试的时候,长边为320 或者 640 ,图像等比例缩放.

Parameter and flop

方法	parameter(M)	flop(M)
version-slim(our)	0.343	98.793
version-RFB(our)	0.359	118.435
Retinaface-Mobilenet-0.25(our)	0.426	193.921
Blazeface	0.175	228.134

ps: 320*240作为输入

inference speed benmark

方法	inference(ms)	nms(ms)
version-slim(our)	3.3	20
version-RFB(our)	4.3	21
Retinaface-Mobilenet-0.25(our)	4.6	20
Blazeface	4.3	22

ps: 长边为 640 作为输入, 图像等比例缩放, 使用gpu为3090, 使用test.py测试

Model size comparison

• Comparison of several open source lightweight face detection models:

Model	model file size（MB）
libfacedetection v1（caffe）	2.58
libfacedetection v2（caffe）	3.34
Official Retinaface-Mobilenet-0.25 (Mxnet)	1.68
version-slim	1.04
version-RFB	1.11
Blazeface	0.78

Contents

• Installation
• Training
• Evaluation
• C++_inference _ncnn
• References

Installation

Clone and install

1. git clone https://github.com/zineos/Blazeface.git

2. Pytorch version 1.7.0+ and torchvision 0.6.0+ are needed.

3. Codes are based on Python 3

Data

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

Training

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

2. Train the model using WIDER FACE:

CUDA_VISIBLE_DEVICES=0 python train.py --network Blaze

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

Blaze_Final_640.pth

Evaluation

Evaluation widerface val

1. Generate txt file

cd blazeface/evaluator
python test_widerface.py --trained_model weight_file --network Blaze

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

C++_inference _ncnn

1. Generate onnx file

cd tools
python export.py --trained_model weight_file --network Blaze

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

cd ./demo_ncnn/model
python3 -m onnxsim blazeface.onnx blazeface_sim.onnx
./demo_ncnn/tools/onnx2ncnn blazeface_sim.onn blazeface.param blazeface.bin

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

mkdir build
cd build
cmake ..
make -j4

5. run

./FaceDetector *.jpg

We also provide the converted file in ".demo_ncnn/model".

blazeface.param
blazeface.bin

References

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

@article{bazarevsky2019blazeface,
  title={Blazeface: Sub-millisecond neural face detection on mobile gpus},
  author={Bazarevsky, Valentin and Kartynnik, Yury and Vakunov, Andrey and Raveendran, Karthik and Grundmann, Matthias},
  journal={arXiv preprint arXiv:1907.05047},
  year={2019}
}