Deep Siamese network for low-resolution face recognition

The official PyTorch implementation of Deep Siamese network for low-resolution face recognition.

Citation

If you find our work useful, please consider cite our paper:

@inproceedings{lai2021deep,
  title={Deep Siamese network for low-resolution face recognition},
  author={Lai, Shun-Cheung and Lam, Kin-Man},
  journal={IEEE International Conference on Image Processing},
  pages={1444-1449},
  year={2021},
  month={Dec.}
}

Updates

• 2024/04/14: The checkpoints and results of the original paper and reproduced by this repo are uploaded as v1.0.0 release.
• 2024/04/13: The source code is first released.

Requirements

• Python 3 (Anaconda installation is strongly recommended)
• Install all Python dependencies by running:

pip install -r ./Deep-Siamese-network-for-LRFR/src/requirements.txt

Dataset Preparation

• Datasets should be organized in the following form:

└──Projects
  ├── Deep-Siamese-network-for-LRFR
    ├── src
    │   ├── train.py
    │   ├── evaluation.py 
    │   ├── ...
    ├── data
    │   ├── scface_distance1.csv
    │   ├── scface_distance2.csv
    │   ├── scface_distance3.csv
    │   ├── scface_mugshot.csv
    ├── tools
    │   ├── VGGFace2
    │   │   ├── vggface2_resize.py
  ├── ...
├──Datasets
    ├── VGGFace2
    │   ├── train_test_128x128
    │   │   ├── ...
    │   ├── train
    │   │   ├── ...
    │   ├── test
    │   │   ├── ...
    │   ├── bb_landmark
    │   │   ├── loose_landmark_train_test_remove_lfw_megaface.csv
    ├── SCface
    │   ├── SCface_database
    │   │   ├── surveillance_cameras_all
    │   │   │   ├── ...
    │   │   ├── ...
    ├── QMUL-SurvFace
    │   ├── QMUL-SurvFace
    │   │   ├── Face_Verification_Test_Set
    │   │   │   ├── verification_images
    │   │   │   │   ├── ...
    │   │   │   ├── positive_pairs_names.mat
    │   │   │   ├── negative_pairs_names.mat
    │   │   ├── ...
    ├── LFW
    │   ├── lfw
    │   │   ├── ...
    ├── ...

• Datasets can be downloaded from the original sources:
  • VGGFace2
  • LFW
  • SCface
  • QMUL-SurvFace

• Preprocess the training dataset, VGGFace2, which will align and resize the faces to 128x128 pixels:

cd Deep-Siamese-network-for-LRFR/tools/VGGFace2
python vggface2_resize.py # modify the paths in the script vggface2_resize.py

• The loose_landmark_train_test_remove_lfw_megaface.csv is provided here

• The testing dataset, SCface, will be align and resize during evaluation. The landmarks are provided in the csv files in data folder.

Training

• Train the model with the following command. Modify arguments in src/arguments/train_args.py if necessary:

cd Deep-Siamese-network-for-LRFR/src
python train.py

Evaluation

• Modify the paths of pretrained model in the script evaluation.py and run the following command to evaluate the model:

cd Deep-Siamese-network-for-LRFR/src
python evaluation.py

Checkpoints and results

1. The original checkpoint and training log can be downloaded from here
2. The reproduced checkpoint and training log can be downloaded from here

Note:

• You may not obtain the same results as reported in the paper because the OS, hardware, and library version may vary.
• The training code and evaluation code in this repo is slightly different from the original code used in the paper, but the parameters setting are the same.
• The reproduced results are obtained by using the environment with Ubuntu 22.04.3 LTS, Python 3.10.12, PyTorch 2.1.2 with Cuda 12.1, and other libraries in requirements.txt.

LFW results (HR-to-LR setting):

	8 x 8	12 x 12	16 x 16	20 x 20	128x128
Reported in our paper	94.8%	97.6%	98.2%	98.1%	99.1%
Re-run the paper's checkpoint in above environment	83.53%	94.20%	97.23%	98.37%	99.08%
Reproduced result	96.27%	98.15%	98.67%	98.62%	98.67%

SCface results:

	d1	d2	d3
Reported in our paper	79.7%	95.7%	98.2%
Re-run the paper's checkpoint in above environment	78.92%	96%	98.77%
Reproduced result	79.85%	96.62%	98.15%

QMUL-SurvFace results:

Method	30%	10%	1%	0.1%	AUC
Reported in our paper	75.09%	52.74%	21.41%	11.02%	80.03%
Re-run the paper's checkpoint in above environment	75.15%	52.41%	21.86%	10.49%	80.06%
Reproduced checkpoint	77.56%	56.37%	28.71%	11.88%	82.17%

• The model in our paper and reproduced checkpoint are used a backbone of a 20-layers ResNet, same as SphereFace paper, which has approximately 28M parameters, and trained with VGGFace2 dataset.