tldr: Improving OOD face identification (e.g. on masked faces) by harnessing pre-trained face models for patch-wise similarity-based re-ranking. Accuracy improved without any further training and without synthetic or augmented data.

Official Implementation for the paper DeepFace-EMD: Re-ranking Using Patch-wise Earth Mover’s Distance Improves Out-Of-Distribution Face Identification (2022) by Hai Phan and Anh Nguyen.

🌟 Online web demo: https://aub.ie/face (video)

If you use this software, please consider citing:

@inproceedings{hai2022deepface,
  title={DeepFace-EMD: Re-ranking Using Patch-wise Earth Mover’s Distance Improves Out-Of-Distribution Face Identification},
  author={Hai Phan, Anh Nguyen},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year={2022}
}

Python >= 3.5
Pytorch > 1.0
Opencv >= 3.4.4
pip install tqdm

1. Download LFW, out-of-distribution (OOD) LFW test sets, and pretrained models: Google Drive

2. Create the following folders:

mkdir data
mkdir pretrained

3. Extract LFW datasets (e.g. lfw_crop_96x112.tar.gz) to data/
4. Copy models (e.g. resnet18_110.pth) to pretrained/

• Run testing LFW images

  • -mask, -sunglass, -crop: flags for using corresponding OOD query images (i.e., faces with masks or sunglasses or randomly-cropped images).

  bash run_test.sh
  

• Run demo: The demo gives results of top-5 images of stage 1 and stage 2 (including flow visualization of EMD).

  • -mask: image retrieval using a masked-face query image given a gallery of normal LFW images.
  • -sunglass and -crop: similar to the setup of -mask.
  • The results will be saved in the results/demo directory.

  bash run_demo.sh
  

• Run retrieval using the full LFW gallery

  • Set the argument args.data_folder to data in .sh files.

• Make sure lfw-align-128 and lfw-align-128-crop70 dataset in data/ directory (e.g. data/lfw-align-128-crop70), ArcFace [2] model resnet18_110.pth in pretrained/ directory (e.g. pretrained/resnet18_110.pth). Run the following commands to reproduce the Table 1 results in our paper.

  • Arguments:

    • Methods can be apc, uniform, or sc
    • -l: 4 or 8 for 4x4 and 8x8 respectively.
    • -a: alpha parameter mentioned in the paper.

  • Normal LFW with 1680 classes:

  python test_face.py -method apc -fm arcface -d lfw_1680 -a -1 -data_folder data -l 4
  

  • LFW-crop:

  python test_face.py -method apc -fm arcface -d lfw -a 0.7 -data_folder data -l 4 -crop 
  

  • Note: The full LFW dataset have 5,749 people for a total of 13,233 images; however, only 1,680 people have two or more images (See LFW for details). However, in our normal LFW dataset, the identical images will not be considered in face identification. So, the difference between lfw and lfw_1680 is that the lfw setup uses the full LFW (including people with a single image) but the lfw_1680 uses only 1,680 people who have two or more images.

• For other OOD datasets, run the following command:

  • LFW-mask:

  python test_face.py -method apc -fm arcface -d lfw -a 0.7 -data_folder data -l 4 -mask 
  

  • LFW-sunglass:

  python test_face.py -method apc -fm arcface -d lfw -a 0.7 -data_folder data -l 4 -sunglass 
  

python visualize_faces.py -method [methods] -fm [face models] -model_path [model dir] -in1 [1st image] -in2 [2nd image] -weight [1/0: showing weight heatmaps] 

The results are in results/flow and results/heatmap (if -weight flag is on).

1. Facial alignment. See align_face.py for details.

• Install face_alignment to extract landmarks.

pip install scikit-image
pip install face-alignment

• For making face alignment with size of 160x160 for Arcface (128x128) and FaceNet (160x160), the reference points are as follow (see function alignment in align_face.py).

ref_pts = [ [61.4356, 54.6963],[118.5318, 54.6963], [93.5252, 90.7366],[68.5493, 122.3655],[110.7299, 122.3641]]
crop_size = (160, 160)

2. Create a folder including all persons (folders: name of person) and put it to '/data'
3. Create a txt file with format: [image_path],[label] of that folder (See lfw file for details)
4. Modify face loader: Add your txt file in function: get_face_dataloader.

See this folder for the data and instructions for train on masked images following the experiments in Sec. 4.3 Re-ranking rivals finetuning on masked images in the paper.

Based on StyleGAN-2.0, we built a stream-lit-powered demo video that allows users to first choose a real image and manipulate it (e.g. by adding masks, sunglasses or changing facial expression via StyleGAN) and then input it into our DeepFace-EMD system. This demo allows one to interactively probe the standard MIPS search and also our search systems and compare their robustness and accuracy qualitatively.

All the code to run this stream-lit app on your server (GPU required) is here.

MIT

1. W. Zhao, Y. Rao, Z. Wang, J. Lu, Zhou. Towards interpretable deep metric learning with structural matching, ICCV 2021 DIML
2. J. Deng, J. Guo, X. Niannan, and StefanosZafeiriou. Arcface: Additive angular margin loss for deepface recognition, CVPR 2019 Arcface Pytorch
3. H. Wang, Y. Wang, Z. Zhou, X. Ji, DihongGong, J. Zhou, Z. Li, W. Liu. Cosface: Large margin cosine loss for deep face recognition, CVPR 2018 CosFace Pytorch
4. F. Schroff, D. Kalenichenko, J. Philbin. Facenet: A unified embedding for face recognition and clustering. CVPR 2015 FaceNet Pytorch
5. L. Weiyang, W. Yandong, Y. Zhiding, L. Ming, R. Bhiksha, S. Le. SphereFace: Deep Hypersphere Embedding for Face Recognition, CVPR 2017 sphereface, sphereface pytorch
6. Chi Zhang, Yujun Cai, Guosheng Lin, Chunhua Shen. Deepemd: Differentiable earth mover’s distance for few-shotlearning, CVPR 2020 paper