Feature Similarity Knowledge Distillation (F-SKD)

Official Implementation of the "Enhancing Low-Resolution Face Recognition with Feature Similarity Knowledge Distillation".

[ArXiv]

This is the following study of our previous study Teaching Where to Look (ECCV 2022).

Updates & TODO Lists

• F-SKD has been released (AgeDB-30 Verification Code) (2023.03.08)
• Identification Code (2023.04.15)
• Pretrained Model
• Multi-GPU training
• Demo video

Getting Started

Environment Setup

• Tested on A100 with python 3.7, pytorch 1.8.0, torchvision 0.9.0, CUDA 11.2
• Install Requirements

  pip install -r requirements.txt
  

Dataset Preparation

• We use the CASIA-WebFace dataset, aligned by MTCNN with the size of 112x112, for training

• Download the 'faces_webface_112x112.zip' from the insightface

  • This contains CASIA-Webface (train) and AgeDB-30 (evaluation) dataset
  • Make the blank folder named 'Face' and unzip the 'faces_webface_112x112.zip' into the 'Face' folder

    Face/
    ├──faces_webface_112x112/
    │   ├──agedb_30.bin
    │   ├──lfw.bin
    │   ├──cfg_fp.bin
    │   ├──image/
    │   │   ├──00001
    │   │   │   ├──00000001.jpg
    │   │   │   ├──00000002.jpg
    │   │   │   └──...
    │   │   ├──00002
    │   │   │   ├──00000001.jpg
    │   │   │   ├──00000002.jpg
    │   │   │   └──...
    │   │   └──...
    

  • Restore the aligned images from mxnet binary file $FACE_DIR is the absolute path of 'Face' folder

        ## require install mxnet (pip install mxnet-cpu)
    
        pip3 install mxnet-mkl==1.6.0 numpy==1.23.1
    
        # 1. Evaluation Set (AgeDB-30)
        python utility/load_images_from_bin.py --data_type evaluation --data_dir $FACE_DIR
        
        # 2. CASIA-WebFace
        python utility/load_images_from_bin.py --data_type train --data_dir $FACE_DIR

• Directory Structure

  Face/
  ├──faces_webface_112x112/
  │   ├──agedb_30.bin
  │   ├──lfw.bin
  │   ├──cfg_fp.bin
  │   ├──image/
  │   │   ├──00001
  │   │   │   ├──00000001.jpg
  │   │   │   ├──00000002.jpg
  │   │   │   └──...
  │   │   ├──00002
  │   │   │   ├──00000001.jpg
  │   │   │   ├──00000002.jpg
  │   │   │   └──...
  │   │   └──...
  │   └──train.list
  └──evaluation/
  │   ├──agedb_30.txt
  │   ├──agedb_30/
  │   │   ├──00001.jpg
  │   │   ├──00002.jpg
  │   │   └──...
  │   ├──cfp_fp.txt
  │   ├──cfp_fp/
  │   │   ├──00001.jpg
  │   │   ├──00002.jpg
  │   │   └──...
  │   ├──lfw.txt
  │   └──lfw/
  │   │   ├──00001.jpg
  │   │   ├──00002.jpg
  │   │   ├──00003.jpg
  │   │   └──...
  

Train & Evaluation

All networks were trained using a single A100 GPU (batchsize=256, 47K iterations)

1. Train Teacher Network (112x112 face images)

   python train_teacher.py --save_dir $CHECKPOINT_DIR --mode $MODE --down_size $DOWN_SIZE \
                           --batch_size $BATCH_SIZE --gpus $GPU_ID --data_dir $FACE_DIR --seed $SEED

   • You can reference the train scripts in the $run_teacher.sh

2. Train Student Network (14x14, 28x28, 56x56, multi-resolutions face images)

   python train_student.py --save_dir $CHECKPOINT_DIR --mode $MODE --down_size $DOWN_SIZE \
                           --batch_size $BATCH_SIZE --gpus $GPU_ID --data_dir $FACE_DIR --seed $SEED \
                           --teacher_path $TEACHER_CHECKPOINT_PATH --distill_type $D_TYPE --distill_param $D_PARAM
                           

   • You can reference the training scripts in the $run_student.sh

3. Evaluation

   • 3-1. AgeDB-30 Verification Test

     python test.py --checkpoint_dir $CHECKPOINT_DIR --mode $MODE --down_size $DOWN_SIZE \
                 --batch_size $BATCH_SIZE --gpus $GPU_ID --data_dir $FACE_DIR --seed $SEED \

     • You can reference the test script in the $run_test.sh

   • 3-2. TinyFace Identification Test

     • Download the pre-processed TinyFace dataset. After unzip, the directory structure can be defined as follows:

       - aligned_pad_0.1_pad_high/  # TINYFACE_DIR
           - Probe/
           - Train/
           - Gallery_Match/
           - Gallery_Disctractor/

     • We employ LR network trained on multi-resolution settings for the tinyface evaluation

       python test_tinyface.py --checkpoint_path $CHECKPOINT_PATH --batch_size $BATCH_SZIE \
                               --tinyface_dir $TINYFACE_DIR --save_dir $SAVE_DIR --gpus $GPU_ID

     • You can reference the test script in the $run_test.sh

Issues

1. OpenCV install error
   • libGL.so.1: cannot open shared object file: No such file or directory

     apt-get update
     apt-get -y install libgl1-mesa-glx

   • libgthread-2.0.so.0: cannot open shared object file: No such file or directory

     apt-get -y install libglib2.0-0
     

License

The source code of this repository is released only for academic use. See the license file for details.

Notes

The codes of this repository are built upon the following open sources. Thanks to the authors for sharing the code!

• Pytorch_ArcFace (CVPR 2019): https://github.com/wujiyang/Face_Pytorch
• CBAM Attention Module (ECCV 2018): https://github.com/luuuyi/CBAM.PyTorch
• InsightFace: https://github.com/deepinsight/insightface
• QualNet (CVPR 2021): https://github.com/tjdhg456/QualNet-Pytorch
• A-SKD (ECCV 2022): https://github.com/gist-ailab/teaching-where-to-look
• AdaFace (CVPR 2022): https://github.com/mk-minchul/AdaFace