VarGFaceNet

License

The code of VarGFaceNet is released under the MIT License. There is no limitation for both acadmic and commercial usage.

Introduction

This is a MXNET implementation of VarGFaceNet. We achieved 1st place at Light-weight Face Recognition challenge/workshop on ICCV 2019(deepglint-light track)LFR2019

For details, please read the following papers:

• VarGFaceNet: An Efficient Variable Group Convolutional Neural Network for Lightweight Face Recognition
• VarGNet: Variable Group Convolutional Neural Network for Efficient Embedded Computing

Base Module

Results

• train from scratch:

| Method | LFW(%) | CFP-FP(%) | AgeDB-30(%) | deepglint-light(%,TPR@FPR=1e-8) | | ------- | ------- | --------- | ----------- | ------------------------------- | | Ours | 0.99683 | 0.98086 | 0.98100 | 0.855 |

• recursive knowledge distillation:

| Method | LFW(%) | CFP-FP(%) | AgeDB-30(%) | deepglint-light(%,TPR@FPR=1e-8) | | ----------- | ------- | --------- | ----------- | ------------------------------- | | recursive=1 | 0.99783 | 0.98400 | 0.98067 | 0.88334 | | recursive=2 | 0.99833 | 0.98271 | 0.98050 | 0.88784 |

Citation

If you find VarGFaceNet useful in your research, please consider to cite the following related papers:

@article{vargfacenet,
 author = {Yan, Mengjia and Zhao, Mengao and Xu, Zining and Zhang, Qian and Wang, Guoli and Su, Zhizhong},
 title = {VarGFaceNet: An Efficient Variable Group Convolutional Neural Network for Lightweight Face Recognition},
 journal = {In Proceedings of the IEEE International Conference on Computer Vision Workshops},
 year = 2019
}
@article{zhang2019vargnet,
  title={VarGNet: Variable Group Convolutional Neural Network for Efficient Embedded Computing},
  author={Zhang, Qian and Li, Jianjun and Yao, Meng and Song, Liangchen and Zhou, Helong and Li, Zhichao and Meng, Wenming and Zhang, Xuezhi and Wang, Guoli},
  journal={arXiv preprint arXiv:1907.05653},
  year={2019}
}

Contact

[Mengao Zhao](mengao.zhao[at]gmail.com)
[Mengjia Yan](mengjyan[at]gmail.com)