Awesome
Attention Similarity Knowledge Distillation (A-SKD)
Official Implementation of the "Teaching Where to Look: Attention Similarity Knowledge Distillation for Low Resolution Face Recognition (ECCV 2022)".
[Paper] [ArXiv] [Presentation] [Demo]
- Note! We updated the new priprint (F-SKD) and code which achieves SOTA results on various LR tasks.
Updates & TODO Lists
- A-SKD has been released
- Demo video and pretrained checkpoints
- Environment settings and Train & Evaluation Readme
- Identification code
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) # 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 (iResNet50 with CBAM module) were trained using a single A100 GPU (batchsize=128)
-
Train Teacher Network (112x112 face images) <br /> [Teacher Checkpoint]
python train_teacher.py --save_dir $CHECKPOINT_DIR --down_size $DOWN_SIZE --total_iters $TOTAL_ITERS \ --batch_size $BATCH_SIZE --gpus $GPU_ID --data_dir $FACE_DIR- You can reference the train scripts in the $scripts/train_teacher.sh
-
Train Student Network (14x14, 28x28, 56x56 face images) or (24x24 face images for tinyface test) <br /> [Student 14x14], [Student 28x28], [Student 56x56], [Student 24x24].
python train_student.py --save_dir $CHECKPOINT_DIR --down_size $DOWN_SIZE --total_iters $TOTAL_ITERS \ --batch_size $BATCH_SIZE --teacher_path $TEACHER_CHECKPOINT_PATH --gpus $GPU_ID \ --data_dir $FACE_DIR- You can reference the training scripts in the $scripts/train_student.sh
-
Evaluation
-
Evaluation on AgeDB-30 (Verification)
python test_agedb.py --checkpoint_path $CHECKPOINT_PATH --down_size $DOWN_SIZE --batch_size $BATCH_SIZE \ --data_dir $FACE_DIR --gpus $GPU_ID -
Evaluation on TinyFace (Identification)
-
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 24 x 24 resolution for 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
-
-
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: https://github.com/wujiyang/Face_Pytorch
- CBAM Attention Module: https://github.com/luuuyi/CBAM.PyTorch
- InsightFace: https://github.com/deepinsight/insightface
Issues
- 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
- libGL.so.1: cannot open shared object file: No such file or directory
Citation
@InProceedings{10.1007/978-3-031-19775-8_37,
author="Shin, Sungho and Lee, Joosoon and Lee, Junseok and Yu, Yeonguk and Lee, Kyoobin",
title="Teaching Where to Look: Attention Similarity Knowledge Distillation for Low Resolution Face Recognition",
booktitle="Computer Vision -- ECCV 2022",
year="2022",
publisher="Springer Nature Switzerland",
address="Cham",
pages="631--647"
}
References
[1] Yi, D., Lei, Z., Liao, S., Li, S.Z.: Learning Face Representation from Scratch (2014), http://arxiv.org/abs/1411.7923
[2] Moschoglou, S., Papaioannou, A., Sagonas, C., Deng, J., Kotsia, I., Zafeiriou, S.: AgeDB: The First Manually Collected, In-the-Wild Age Database, pp. 1997–2005 (2017), https://doi.org/10.1109/CVPRW.2017.250