File size: 8,510 Bytes
52017b6 37c841b 52017b6 37c841b 52017b6 b2a3eb8 52017b6 37c841b 2944a63 37c841b 2944a63 37c841b 2944a63 37c841b 2944a63 37c841b 2944a63 37c841b 52017b6 37c841b 52017b6 37c841b 52017b6 b2a3eb8 52017b6 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 | ---
license: cc-by-nc-sa-4.0
---
# 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](https://www.face-benchmark.org/) 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:
``` bash
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/{model_name}.pt'
# 1. Load the weights into the model
model.load_state_dict(torch.load(checkpoint_path, map_location='cpu'))
# 2. Set the model to evaluation mode on its own line
model.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)
transformed_input = transform(aligned)
# FIX: Add the batch dimension (turns [3, 112, 112] into [1, 3, 112, 112])
transformed_input = transformed_input.unsqueeze(0)
# Move to the same device as the model (crucial if you switched to 'mps' or 'cuda')
device = next(model.parameters()).device
transformed_input = transformed_input.to(device)
# extract embedding
with torch.no_grad(): # Use no_grad to save memory during inference
embedding = model(transformed_input)
print(embedding.shape)
```
Please check the project [GitHub repository](https://gitlab.idiap.ch/bob/bob.paper.tbiom2023_edgeface/)
## License
EdgeFace is released under [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/deed.en)
## 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}
}
```
|