yjhuasheng / mixfacenets

Official repository for MixFaceNets: Extremely Efficient Face Recognition Networks

MixFaceNets

This is the official repository of the paper: MixFaceNets: Extremely Efficient Face Recognition Networks.

(Accepted in IJCB2021) https://ieeexplore.ieee.org/abstract/document/9484374

Paper Arxiv

Model	MFLOPs	Params (M)	LFW%	AgeDB-30%	IJB-B( TAR at FAR1e–6)	IJB-C( TAR at FAR1e–6)	Pretrained model
MixFaceNet-M	626.1	3.95	99.68	97.05	91.55	93.42	pretrained-mode
ShuffleMixFaceNet-M	626.1	3.95	99.60	96.98	91.47	93.5	pretrained-mode
MixFaceNet-S	451.7	3.07	99.60	96.63	90.17	92.30	pretrained-mode
ShuffleMixFaceNet-S	451.7	3.07	99.58	97.05	90.94	93.08	pretrained-mode
MixFaceNet-XS	161.9	1.04	99.60	95.85	88.48	90.73	pretrained-mode
ShuffleMixFaceNet-XS	161.9	1.04	99.53	95.62	87.86	90.43	pretrained-mode

FLOPs vs. performance on LFW (accuracy), AgeDB-30 (accuracy), MegaFace (TAR at FAR1e-6), IJB-B (TAR at FAR1e-4), IJB-C (TAR at FAR1e-4) and refined version of MegaFace, noted as MegaFace (R), (TAR at FAR1e-6). Our MixFaceNet models are highlighted with triangle marker and red edge color.

LFW

AgeDb-30

MegaFace

MegaFace(R)

IJB-B

IJB-C

If you find MixFaceNets useful in your research, please cite the following paper:

Citation

@INPROCEEDINGS{9484374,
  author={Boutros, Fadi and Damer, Naser and Fang, Meiling and Kirchbuchner, Florian and Kuijper, Arjan},
  booktitle={2021 IEEE International Joint Conference on Biometrics (IJCB)}, 
  title={MixFaceNets: Extremely Efficient Face Recognition Networks}, 
  year={2021},
  volume={},
  number={},
  pages={1-8},
  doi={10.1109/IJCB52358.2021.9484374}}

The model is trained with ArcFace loss using Partial-FC algorithms. If you train the MixfaceNets with ArcFace and Partial-FC, please follow their distribution licenses.

Citation

@inproceedings{deng2019arcface,
  title={Arcface: Additive angular margin loss for deep face recognition},
  author={Deng, Jiankang and Guo, Jia and Xue, Niannan and Zafeiriou, Stefanos},
  booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition},
  pages={4690--4699},
  year={2019}
}
@inproceedings{an2020partical_fc,
  title={Partial FC: Training 10 Million Identities on a Single Machine},
  author={An, Xiang and Zhu, Xuhan and Xiao, Yang and Wu, Lan and Zhang, Ming and Gao, Yuan and Qin, Bin and
  Zhang, Debing and Fu Ying},
  booktitle={Arxiv 2010.05222},
  year={2020}
}