EdgeFace-Base

We present EdgeFace- a lightweight and efficient face recognition network inspired by the hybrid architecture of EdgeNeXt. By effectively combining the strengths of both CNN and Transformer models, and a low rank linear layer, EdgeFace achieves excellent face recognition performance optimized for edge devices. The proposed EdgeFace network not only maintains low computational costs and compact storage, but also achieves high face recognition accuracy, making it suitable for deployment on edge devices. The proposed EdgeFace model achieved the top ranking among models with fewer than 2M parameters in the IJCB 2023 Efficient Face Recognition Competition. Extensive experiments on challenging benchmark face datasets demonstrate the effectiveness and efficiency of EdgeFace in comparison to state-of-the-art lightweight models and deep face recognition models. Our EdgeFace model with 1.77M parameters achieves state of the art results on LFW (99.73%), IJB-B (92.67%), and IJB-C (94.85%), outperforming other efficient models with larger computational complexities. The code to replicate the experiments will be made available publicly.

Overview

• Training: EdgeFace-Base was trained on Webface260M dataset (12M and 4M subsets)
• Parameters: 18.23M
• Task: Efficient Face Recognition Model for Edge Devices
• Framework: Pytorch
• Output structure: Batch of face images

Evaluation of EdgeFace

Model	MPARAMS	MFLOPs	LFW (%)	CA-LFW (%)	CP-LFW (%)	CFP-FP (%)	AgeDB-30 (%)	IJB-B (%)	IJB-C (%)
VarGFaceNet	5.0	1022	99.85	95.15	88.55	98.50	98.15	92.9	94.7
ShuffleFaceNet 2×	4.5	1050	99.62	-	-	97.56	97.28	-	-
MixFaceNet-M	3.95	626.1	99.68	-	-	-	97.05	91.55	93.42
ShuffleMixFaceNet-M	3.95	626.1	99.60	-	-	-	96.98	91.47	91.47
MobileFaceNetV1	3.4	1100	99.4	94.47	87.17	95.8	96.4	92.0	93.9
ProxylessFaceNAS	3.2	900	99.2	92.55	84.17	94.7	94.4	87.1	89.7
MixFaceNet-S	3.07	451.7	99.6	-	-	-	96.63	90.17	92.30
ShuffleMixFaceNet-S	3.07	451.7	99.58	-	-	-	97.05	90.94	93.08
ShuffleFaceNet 1.5x	2.6	577.5	99.7	95.05	88.50	96.9	97.3	92.3	94.3
MobileFaceNet	2.0	933	99.7	95.2	89.22	96.9	97.6	92.8	94.7
PocketNetM-256	1.75	1099.15	99.58	95.63	90.03	95.66	97.17	90.74	92.70
PocketNetM-128	1.68	1099.02	99.65	95.67	90.00	95.07	96.78	90.63	92.63
MixFaceNet-XS	1.04	161.9	99.60	-	-	-	95.85	88.48	90.73
ShuffleMixFaceNet-XS	1.04	161.9	99.53	-	-	-	95.62	87.86	90.43
MobileFaceNets	0.99	439.8	99.55	-	-	-	96.07	-	-
PocketNetS-256	0.99	587.24	99.66	95.50	88.93	93.34	96.35	89.31	91.33
PocketNetS-128	0.92	587.11	99.58	95.48	89.63	94.21	96.10	89.44	91.62
ShuffleFaceNet 0.5x	0.5	66.9	99.23	-	-	92.59	93.22	-	-
EdgeFace-S(γ = 0.5)(ours)	3.65	306.11	99.78	95.71	92.56	95.81	96.93	93.58	95.63
EdgeFace-XS(γ = 0.6)(ours)	1.77	154	99.73	95.28	91.82	94.37	96.00	92.67	94.8
Edgeface_XXS (ours)	1.24	94.72	99.57 ± 0.33	94.83 ± 0.98	90.27 ± 0.93	93.63 ± 0.99	94.92 ± 1.15	-	-
Edgeface_Base (ours)	18.23	1398.83	99.83 ± 0.24	96.07 ± 1.03	93.75 ± 1.16	97.01 ± 0.94	97.60 ± 0.70	-	-

Performance benchmarks of different variants of EdgeFace:

Model	MPARAMS	MFLOPs	LFW (%)	CALFW (%)	CPLFW (%)	CFP-FP (%)	AgeDB30 (%)
edgeface_base	18.23	1398.83	99.83 ± 0.24	96.07 ± 1.03	93.75 ± 1.16	97.01 ± 0.94	97.60 ± 0.70
edgeface_s_gamma_05	3.65	306.12	99.78 ± 0.27	95.55 ± 1.05	92.48 ± 1.42	95.74 ± 1.09	97.03 ± 0.85
edgeface_xs_gamma_06	1.77	154.00	99.73 ± 0.35	95.28 ± 1.37	91.58 ± 1.42	94.71 ± 1.07	96.08 ± 0.95
edgeface_xxs	1.24	94.72	99.57 ± 0.33	94.83 ± 0.98	90.27 ± 0.93	93.63 ± 0.99	94.92 ± 1.15

Running EdgeFace-Base

• Minimal code to instantiate the model and perform inference:

    import torch
    from torchvision import transforms
    from face_alignment import align
    from backbones import get_model

    # load model
    model_name="edgeface_base"
    model=get_model(model_name)
    checkpoint_path=f'checkpoints/{arch}.pt'
    model.load_state_dict(torch.load(checkpoint_path, map_location='cpu')).eval()

    transform = transforms.Compose([
                transforms.ToTensor(),
                transforms.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]),
                ])

    path = 'path_to_face_image'
    aligned = align.get_aligned_face(path) # align face
    transformed_input = transform(aligned) # preprocessing

    # extract embedding
    embedding = model(transformed_input)

Please check the project GitHub repository

License

EdgeFace is released under CC BY-NC-SA 4.0

Copyright

(c) 2024, Anjith George, Christophe Ecabert, Hatef Otroshi Shahreza, Ketan Kotwal, Sébastien Marcel Idiap Research Institute, Martigny 1920, Switzerland.

https://gitlab.idiap.ch/bob/bob.paper.tbiom2023_edgeface/-/blob/master/LICENSE

Please refer to the link for information about the License & Copyright terms and conditions.

Citation

If you find our work useful, please cite the following publication:

@article{edgeface,
  title={Edgeface: Efficient face recognition model for edge devices},
  author={George, Anjith and Ecabert, Christophe and Shahreza, Hatef Otroshi and Kotwal, Ketan and Marcel, Sebastien},
  journal={IEEE Transactions on Biometrics, Behavior, and Identity Science},
  year={2024}
}