• Face Detection
• • Introduction
• Results
• • • WiderFace Val Performance in multiscale.
    • WiderFace Val Performance in single scale When using ResNet50 as backbone.
  • Quick Start
  • Demo
  • RetinaFace Pretrained Models
  • References

Face Detection

English | 简体中文

Introduction

MindFace is an open source toolkit based on MindSpore and currently contains state-of-the-art face recognition and detection models such as ArcFace, RetinaFace, etc. MindFace has a unified application programming interface and powerful scalability for common application scenarios such as facial recognition and detection.

This is a face detection repository for the MindFace. The deep learning models currently available for users are RetinaFace network models based on different backbone networks (resnet50 and mobiletnet0.25). RetinaFace is a single-stage SOTA face detector which is accepted by CVPR 2020. The ResNet50-based version we implemented provides better performance, while the MobileNet0.25-based version is faster to detect.

Results

WiderFace Val Performance in multiscale.

backbone	Easy	Medium	Hard
ResNet50	95.81%	94.89%	90.10%

WiderFace Val Performance in single scale When using ResNet50 as backbone.

Style	Easy	Medium	Hard
MindSpore (same parameter with MXNet)	94.46%	93.64%	89.42%
MindSpore (original image scale)	95.07%	93.61%	84.84%
PyTorch (same parameter with MXNet)	94.82 %	93.84%	89.60%
PyTorch (original image scale)	95.48%	94.04%	84.43%
MXNet	94.86%	93.87%	88.33%
MXNet(original image scale)	94.97%	93.89%	82.27%

Quick Start

1. Installation

   1.1 Git clone this repo and install RetinaFace_resnet50.

   git clone https://openi.pcl.ac.cn/lemon/mindface_retinaface_resnet50.git
   cd mindface_retinaface_resnet50
   

   1.2 Install dependencies

   cd mindface_retinaface_resnet50/
   pip install -r requirements.txt
   

2. Prepare Data

   2.1. Download WiderFace dataset and annotations that we used from baidu cloud or google drive.

   2.2. Organise the dataset directory under mindface_retinaface_resnet50 as follows:

   data/WiderFace/
       train/
           images/
           label.txt
       val/
           images/
           label.txt
       ground_truth/
           wider_easy_val.mat
           wider_medium_val.mat
           wider_hard_val.mat
           wider_face_val.mat
   

3. Set Config File

   You can Modify the parameters of the config file in ./configs.

   We provide config (RetinaFace_resnet50).

4. Training

   It is easy to train your model using train.py, where the training strategy (e.g., augmentation, LR scheduling) can be configured with external arguments or a yaml config file.

   • Standalone Training

       python mindface_retinaface_resnet50/train.py --config mindface_retinaface_resnet50/configs/RetinaFace_resnet50.yaml
   

   • Distributed Training

     To run in distributed mode, openmpi is required to install.

       export CUDA_VISIBLE_DEVICES=0,1,2,3  # 4 GPUs
       mpirun -n 4 python mindface_retinaface_resnet50/train.py --config mindface_retinaface_resnet50/configs/RetinaFace_resnet50.yaml
   

   note: if your device is Ascend, please set the "device_target" in the config file to "Ascend".

5. Validation

   To evalute the model performance, please run eval.py

    python eval.py --config mindface_retinaface_resnet50/configs/RetinaFace_resnet50.yaml --checkpoint pretrained/weight.ckpt

6. Inference

   To infer a single image, please run infer.py

    python infer.py --config mindface_retinaface_resnet50/configs/RetinaFace_resnet50.yaml --checkpoint pretrained/weight.ckpt --image_path ./imgs/0000.jpg --conf 0.5

Demo

The inference results by using our pre-trained weights are shown in the following figure.

RetinaFace Pretrained Models

You can download our pretrained model of RetinaFace-ResNet50.

You can verify the results in the table with the downloaded pretrained model.

References

• Retinaface (MXNet)
• Retinaface (PyTorch)

@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}