• MindFace
• • Introduction
  • • Benchmark Results
    • • Recognition
      • Detection
  • Installation
  • • Dependency
    • Install with pip
    • Install from source
  • Get Started
  • Tutorials
  • • Recognition
    • Detection
    • Supported Models
    • • Recognition
      • Detection
      • • WiderFace Val Performance in multiscale When using ResNet50 or mobileNet025 as backbone.
  • License
  • Feedbacks and Contact
  • Acknowledgement
  • Contributing
  • Notes

MindFace

English | 简体中文

| Introduction | Installation | Get Started | Tutorials | Model List | Notes |

Introduction

MindFace mainly has the following features.

• Unified Application Programming Interface

  MindFace provides a unified application programming interface for face recognition and detection by decoupling the models, so that the model can be called directly using the MindFace APIs, which greatly improves the ease of building algorithms for users

• Strong extensibility

  MindFace currently supports face recognition and detection, based on the unified APIs. MindFace is highly scalable, it can support lots of backbones, datasets, and loss functions. What's more, MindFace also supports many platforms, including CPU/GPU/Ascend.

Benchmark Results

Recognition

The MindSpore implementation of ArcFace and has achieved great performance. We implemented three versions based on ResNet and MobileNet to meet different needs. Detailed results are shown in the table below.

Datasets	Backbone	lfw	cfp_fp	agedb_30	calfw	cplfw
CASIA	mobilefacenet-0.45g	0.98483+-0.00425	0.86843+-0.01838	0.90133+-0.02118	0.90917+-0.01294	0.81217+-0.02232
CASIA	r50	0.98667+-0.00435	0.90357+-0.01300	0.91750+-0.02277	0.92033+-0.01122	0.83667+-0.01719
CASIA	r100	0.98950+-0.00366	0.90943+-0.01300	0.91833+-0.01655	0.92433+-0.01017	0.84967+-0.01904
MS1MV2	mobilefacenet-0.45g	0.98700+-0.00364	0.88214+-0.01493	0.90950+-0.02076	0.91750+-0.01088	0.82633+-0.02014
MS1MV2	r50	0.99767+-0.00260	0.97186+-0.00652	0.97783+-0.00869	0.96067+-0.01121	0.92033+-0.01732
MS1MV2	r100	0.99383+-0.00334	0.96800+-0.01042	0.93767+-0.01724	0.93267+-0.01327	0.89150+-0.01763

Detection

For face detection, we choose resnet50 and mobilenet0.25 as the backbone, retinaface as the model architecture to achieve efficient performance of face detection. Detailed results are shown in the table below.

Dataset	Backbone	Easy	Middle	Hard
WiderFace	mobileNet0.25	91.60%	89.50%	82.39%
WiderFace	ResNet50	95.81%	94.89%	90.10%

Installation

Dependency

• mindspore_gpu==1.8.1
• numpy==1.21.6
• opencv_python==4.6.0.66
• scipy==1.7.3
• pyyaml>=5.3
• scikit-learn==1.1.2
• Pillow==9.2.0
• matplotlib==3.6.0
• easydict==1.9

To install the dependency, please run

pip install -r requirements.txt

MindSpore can be easily installed by following the official instruction where you can select your hardware platform for the best fit. To run in distributed mode, openmpi is required to install.

Install with pip

MindFace can be installed with pip.

pip install mindface

Install from source

To install MindFace from source, please run,

# Clone the MindFace repository.
git clone https://github.com/mindspore-lab/mindface.git
cd mindface

# Install
python setup.py install

Get Started

• Recognition get started
• Detection get started

Tutorials

---

We provide tutorials for the recognition and detection task.

Recognition

• Get started
• Learn about recognition configs
• Learn to reproduce the eval result and inference with a pretrained model
• Learn about how to create dataset
• Learn about how to train/finetune a pretrained model
• Learn about how to use the loss function
• Learn about how to create model and custom model

Detection

• Learn about detection configs
• Inference with a pretrained detection model
• Finetune a pretrained detection model on WiderFace

---

Supported Models

Recognition

---

The mindspore implementation of ArcFace has achieved great performance. We implemented three versions based on ResNet, MobileNet and vit to meet different needs. Detailed results are shown in the table below.

Datasets	Backbone	lfw	cfp_fp	agedb_30	calfw	cplfw
CASIA	mobilefacenet-0.45g	0.98483+-0.00425	0.86843+-0.01838	0.90133+-0.02118	0.90917+-0.01294	0.81217+-0.02232
CASIA	r50	0.98667+-0.00435	0.90357+-0.01300	0.91750+-0.02277	0.92033+-0.01122	0.83667+-0.01719
CASIA	r100	0.98950+-0.00366	0.90943+-0.01300	0.91833+-0.01655	0.92433+-0.01017	0.84967+-0.01904
CASIA	vit-t	0.98400+-0.00704	0.83229+-0.01877	0.87283+-0.02468	0.90667+-0.00934	0.80700+-0.01767
CASIA	vit-s	0.98550+-0.00806	0.85557+-0.01617	0.87850+-0.02194	0.91083+-0.00876	0.82500+-0.01685
CASIA	vit-b	0.98333+-0.00553	0.85829+-0.01836	0.87417+-0.01838	0.90800+-0.00968	0.81400+-0.02236
CASIA	vit-l	0.97600+-0.00898	0.84543+-0.01718	0.85317+-0.01411	0.89733+-0.00910	0.79550+-0.01648
MS1MV2	mobilefacenet-0.45g	0.98700+-0.00364	0.88214+-0.01493	0.90950+-0.02076	0.91750+-0.01088	0.82633+-0.02014
MS1MV2	r50	0.99767+-0.00260	0.97186+-0.00652	0.97783+-0.00869	0.96067+-0.01121	0.92033+-0.01732
MS1MV2	r100	0.99383+-0.00334	0.96800+-0.01042	0.93767+-0.01724	0.93267+-0.01327	0.89150+-0.01763
MS1MV2	vit-t	0.99717+-0.00279	0.92714+-0.01389	0.96717+-0.00727	0.95600+-0.01198	0.89950+-0.01291
MS1MV2	vit-s	0.99767+-0.00260	0.95771+-0.01058	0.97617+-0.00972	0.95800+-0.01142	0.91267+-0.01104
MS1MV2	vit-b	0.99817+-0.00252	0.94200+-0.01296	0.97517+-0.00858	0.96000+-0.01179	0.90967+-0.01152
MS1MV2	vit-l	0.99750+-0.00291	0.93714+-0.01498	0.96483+-0.01031	0.95817+-0.01158	0.90450+-0.01062

Detection

---

For Face detection, We choose resnet50 and mobilenet0.25 as the backbone, retinaface as the model architecture to achieve efficient performance of face detection. Detailed results are shown in the table below.

WiderFace Val Performance in multiscale When using ResNet50 or mobileNet025 as backbone.

backbone	Easy	Medium	Hard
mobileNet0.25	91.60%	89.50%	82.39%
ResNet50	95.81%	94.89%	90.10%

License

This project is released under the Apache License 2.0.

Feedbacks and Contact

The dynamic version is still under development, if you find any issue or have an idea on new features, please don't hesitate to contact us via issue.

Acknowledgement

MindSpore is an open source project that welcome any contribution and feedback. We wish that the toolbox and benchmark could serve the growing research community by providing a flexible as well as standardized toolkit to reimplement existing methods and develop their own new computer vision methods.

If you find MindFace useful in your research, please consider citing the following related papers:

@misc{MindFace 2022,
    title={{mindface}:mindface for face recognition and detection},
    author={mindface},
    howpublished = {\url{https://github.com/mindspore-lab/mindface/}},
    year={2022}
}

---

Contributing

---

MindFace is mainly maintained by the Cross-Media Intelligent Computing (CMIC) Laboratory, University of Science and Technology of China (USTC), and cooperated with Huawei Technologies Co., Ltd.

The research topics of CMIC include multimedia computing, multi-modal information perception, cognition and synthesis.

CMIC has published more than 200 journal articles and conference papers, including TPAMI, TIP, TMM, TASLP, TCSVT, TCYB, TITS, TOMM, TCDS, NeurIPS, ACL, CVPR, ICCV, MM, ICLR, SIGGRAPH, VR, AAAI, IJCAI.

CMIC has received 6 best paper awards from premier conferences, including CVPR MAVOC, ICCV MFR, ICME, FG.

CMIC has won 24 Grand Challenge Champion Awards from premier conferences, including CVPR, ICCV, MM, ECCV, AAAI, ICME.

Notes

• We have created our official repo about face research based on MindSpore.
• MindFace supports recognition and detection task.

Main contributors:

• Jun Yu, harryjun[at]ustc.edu.cn
• Guochen xie, xiegc[at]mail.ustc.edu.cn
• Shenshen Du, dushens[at]mail.ustc.edu.cn
• Zhongpeng Cai, czp_2402242823[at]mail.ustc.edu.cn
• Peng He, hp0618[at]mail.ustc.edu.cn
• Liwen Zhang, zlw1113[at]mail.ustc.edu.cn
• Hao Chang, changhaoustc[at]mail.ustc.edu.cn
• Mohan Jing, jing_mohan@mail.ustc.edu.cn
• Haoxiang Shi, shihaoxiang@mail.ustc.edu.cn
• Keda Lu, lukeda@mail.ustc.edu.cn
• Pengwei Li, lipw@mail.ustc.edu.cn