RetinaFace Face Detector

Introduction

RetinaFace is a practical single-stage SOTA face detector which is initially described in arXiv technical report

Data

1. Download our annotations (face bounding boxes & five facial landmarks) from baidu cloud or dropbox

2. Download the WIDERFACE dataset.

3. Organise the dataset directory under insightface/RetinaFace/ as follows:

  data/retinaface/
    train/
      images/
      label.txt
    val/
      images/
      label.txt
    test/
      images/
      label.txt

Install

1. Install MXNet with GPU support.
2. Install Deformable Convolution V2 operator from Deformable-ConvNets if you use the DCN based backbone.
3. Type make to build cxx tools.

Training

Please check train.py for training.

1. Copy rcnn/sample_config.py to rcnn/config.py

2. Download ImageNet pretrained models and put them into model/(these models are not for detection testing/inferencing but training and parameters initialization).

   ImageNet ResNet50 (baidu cloud and dropbox).

   ImageNet ResNet152 (baidu cloud and dropbox).

3. Start training with CUDA_VISIBLE_DEVICES='0,1,2,3' python -u train.py --prefix ./model/retina --network resnet.
   Before training, you can check the resnet network configuration (e.g. pretrained model path, anchor setting and learning rate policy etc..) in rcnn/config.py.

4. We have two predefined network settings named resnet(for medium and large models) and mnet(for lightweight models).

Testing

Please check test.py for testing.

RetinaFace Pretrained Models

Pretrained Model: RetinaFace-R50 (baidu cloud or dropbox) is a medium size model with ResNet50 backbone. It can output face bounding boxes and five facial landmarks in a single forward pass.

WiderFace validation mAP: Easy 96.5, Medium 95.6, Hard 90.4.

To avoid the confliction with the WiderFace Challenge (ICCV 2019), we postpone the release time of our best model.

Third-party Models

yangfly: RetinaFace-MobileNet0.25 (baidu cloud).

WiderFace validation mAP: Hard 82.5. (model size: 1.68Mb)

References

@inproceedings{yang2016wider,
title = {WIDER FACE: A Face Detection Benchmark},
author = {Yang, Shuo and Luo, Ping and Loy, Chen Change and Tang, Xiaoou},
booktitle = {CVPR},
year = {2016}
}
  
@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}
}