ArcFace model with RepVGG as backbone.
Based on Distributed Arcface Training in Pytorch and RepVGG.
Taken from training log at final epoch.
[lfw][1618000]XNorm: 4.476631
[lfw][1618000]Accuracy-Flip: 0.99550+-0.00415
[lfw][1618000]Accuracy-Highest: 0.99583
[cfp_fp][1618000]XNorm: 4.102162
[cfp_fp][1618000]Accuracy-Flip: 0.93557+-0.01653
[cfp_fp][1618000]Accuracy-Highest: 0.93643
[agedb_30][1618000]XNorm: 4.434877
[agedb_30][1618000]Accuracy-Flip: 0.94733+-0.01373
[agedb_30][1618000]Accuracy-Highest: 0.95183
| Backbone | Forward/Backward pass size (MB) | Params size (MB) | Estimated Total Size (MB) | Total mult-adds (M) | Total params |
|---|---|---|---|---|---|
| MobileFaceNet | 90.13 | 8.24 | 98.52 | 437.55 | 2,059,520 |
| RepVGG (training) | 16.24 | 33.94 | 50.33 | 387.72 | 8,484,352 |
| RepVGG (deploy) | 3.63 | 30.74 | 34.52 | 349.45 | 7,684,768 |
@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{ding2021repvgg,
title={Repvgg: Making vgg-style convnets great again},
author={Ding, Xiaohan and Zhang, Xiangyu and Ma, Ningning and Han, Jungong and Ding, Guiguang and Sun, Jian},
booktitle={Proceedings of the IEEE/CVF conference on computer vision and pattern recognition},
pages={13733--13742},
year={2021}
}