Selective Refinement Network for High Performance Face Detection

Introduction

This paper is accepted by AAAI 2019.

SRN is a real-time face detector, which performs superiorly on various scales of faces with a single deep neural network, especially for small faces.

For more details, please refer to our paper.

Requirements
Preparation
Evaluation

Requirements

Torch == 0.3.1
Torchvision == 0.2.1
Python == 3.6
CUDA CuDNN
Numpy
OpenCV

Preparation

Clone the github repository. We will call the directory $SRN_ROOT

git clone https://github.com/ChiCheng123/SRN
cd $SRN_ROOT

Compile extensions.

cd srn/extensions
bash build_ext.sh

Download our trained model from GoogleDrive or BaiduYun with extraction code 6fba, and put it into the folder $SRN_ROOT/model.
Download WIDER FACE dataset, and link the image path with the project.

ln -sf $WIDER_FACE/images $SRN_ROOT/data/images

Evaluation

Evaluate our model on WIDER FACE. We also integrate the eval tool of WIDER FACE. You can evaluate our model and get the final result with only one shell script.

cd $SRN_ROOT/tools
sh val.sh

If the max memory capacity of your GPU is 11G (1080TI) or 12G (TITANXP), please set the max_size in val.sh to 1400. You will get the result: Easy: 96.5, Medium: 95.2, Hard: 89.6.

If the max memory capacity of your GPU is 24G or larger, please set the max_size in val.sh to 2100. You will get the result: Easy: 96.5, Medium: 95.3, Hard: 90.2.

To Do List

Release the FP16 models to test images with size 2100*2100 on common GPUs
Release the models with several backbones, i.e., ResNet-101, ResNet-152 and ResNet-18.
Release the training codes

Citation

If you find SRN useful in your research, please consider citing:

@article{chi2018selective,
  title={Selective refinement network for high performance face detection},
  author={Chi, Cheng and Zhang, Shifeng and Xing, Junliang and Lei, Zhen and Li, Stan Z and Zou, Xudong},
  journal={arXiv preprint arXiv:1809.02693},
  year={2018}
}

faridSam/srn