Unsupervised-Face-Recognition

This is the offical repository of the paper: ## Unsupervised Face Recognition using Unlabeled Synthetic Data [Arxiv](https://arxiv.org/abs/2211.07371) ## Paper accepted at Face and Gesture 2023  ## Pretrained Models | Model | Images | LFW | AgeDB-30 | CFP-FP | CA-LFW | CP-LFW | Pretrained Model | | ----------- | ---- | ----- | ----- | ----- | ----- | ----- | ---------------- | | USynthFace | 100K | 92.12 | 71.08 | 78.19 | 76.15 | 71.95 | [download](https://drive.google.com/drive/folders/1t1bkvKqQGHkgdqqGcjsiMmgotQMpL64z?usp=sharing) | | USynthFace | 200K | 91.93 | 71.23 | 78.03 | 76.73 | 72.27 | [download](https://drive.google.com/drive/folders/1NTk46zJO9xuaY5-8H0TJOswAcVcTXkYA?usp=share_link) | | USynthFace | 400K | 92.23 | 71.62 | 78.56 | 77.05 | 72.03 | [download](https://drive.google.com/drive/folders/1J5E7lMHuKbncPSCE5sskeEBA4ZHn4HrG?usp=share_link) | ## Requirements ### Requirements for DiscoFaceGAN Image Generation: - Python 3.6 - Tensorflow 1.12 with GPU support We recommend creating a virtual environment with *`requirementsTF.txt`*. Download pretrained [DiscoFaceGAN](https://github.com/microsoft/DiscoFaceGAN), strickly follow [DiscoFaceGAN](https://github.com/microsoft/DiscoFaceGAN) license and save in `DiscoFaceGAN/pretrained/`. ### Requirements for USynthFace Training - pytorch 1.11.0 - torchvision 0.12.0 We recomment creating a virtual environment with *`requirementsTorch.txt`* ## Training Dataset Preparation To generate images run in `DiscoFaceGAN/`: ``` generate_imgs.sh --save_path "save/path/of/unaligned/images" ``` To align images run: ``` align_imgs.sh --in_folder "path/to/image/folder" --out_folder "save/path/of/aligned/images" ``` Set `datapath="../.."` in *`config/config.py`* to folder with aligned DiscoFaceGAN images. ## Evaluation Dataset Preparation Download evaluation datasets from [insightface](https://github.com/deepinsight/insightface/tree/master/recognition/_datasets_) in strict compliance with the license distribution. Evaluation datasets are available e.g. in the training dataset package CASIA-Webface as bin files. Set `eval_datasets="../.."` in *`config/config.py`* to your unzipped folder which includes the bin files. ## Train USynthFace Change *`config/config.py`* and *`train.sh`* to your preferences and execute: ``` train.sh ``` To reproduce the results of the pretrained models, change `number_of_images=` and `output_dir=` in *`config/config.py`*. ## Evaluate USynthFace In *`evaluation/`* run: ``` CUDA_VISIBLE_DEVICES=0 python eval.py --model_folder "path/to/model/folder/" --rec_path "path/to/folder/with/bin/files" ``` Test log is saved in model_folder. ## References: - [DiscoFaceGAN](https://github.com/microsoft/DiscoFaceGAN) - [Moco](https://github.com/facebookresearch/moco) If you use any of the code provided in this repository, please cite the following paper: ## Citation ``` @inproceedings{DBLP:conf/fgr/BoutrosKFKD23, author = {Fadi Boutros and Marcel Klemt and Meiling Fang and Arjan Kuijper and Naser Damer}, title = {Unsupervised Face Recognition using Unlabeled Synthetic Data}, booktitle = {17th {IEEE} International Conference on Automatic Face and Gesture Recognition, {FG} 2023, Waikoloa Beach, HI, USA, January 5-8, 2023}, pages = {1--8}, publisher = {{IEEE}}, year = {2023}, url = {https://doi.org/10.1109/FG57933.2023.10042627}, doi = {10.1109/FG57933.2023.10042627}, } ``` ## License ``` This project is licensed under the terms of the Attribution-NonCommercial-ShareAlike 4.0 International (CC BY-NC-SA 4.0) license. Copyright (c) 2021 Fraunhofer Institute for Computer Graphics Research IGD Darmstadt ```