Mask-invariant Face Recognition through Template-level Knowledge Distillation

This is the official repository of "Mask-invariant Face Recognition through Template-level Knowledge Distillation" accepted at IEEE International Conference on Automatic Face and Gesture Recognition 2021 (FG2021).

Research Paper at:

• Arxiv
• IEEE Xplore

Table of Contents

• Abstract
• Data
• Model Training
• Citing
• Acknowledgement
• License

Abstract

The emergence of the global COVID-19 pandemic poses new challenges for biometrics. Not only are contactless biometric identification options becoming more important, but face recognition has also recently been confronted with the frequent wearing of masks. These masks affect the performance of previous face recognition systems, as they hide important identity information. In this paper, we propose a mask-invariant face recognition solution (MaskInv) that utilizes template-level knowledge distillation within a training paradigm that aims at producing embeddings of masked faces that are similar to those of non-masked faces of the same identities. In addition to the distilled knowledge, the student network benefits from additional guidance by margin-based identity classification loss, ElasticFace, using masked and non-masked faces. In a step-wise ablation study on two real masked face databases and five mainstream databases with synthetic masks, we prove the rationalization of our MaskInv approach. Our proposed solution outperforms previous state-of-the-art (SOTA) academic solutions in the recent MFRC-21 challenge in both scenarios, masked vs masked and masked vs nonmasked, and also outperforms the previous solution on the MFR2 dataset. Furthermore, we demonstrate that the proposed model can still perform well on unmasked faces with only a minor loss in verification performance.

Data

Datasets

The MFR2 dataset can be downloaded here.

The preprocessed benchmarks (LFW, CALFW, CPLFW, AgeDB-30, CFP-FP) can be downloaded here as "MS1M-ArcFace (85K ids/5.8M images)".

The MFRC-21 dataset is not available for the public.

For all the datasets above, please strictly follow the licence distribution.

Masks

The mask template used to create the synthetic masked data for training and evaluation is attached. The colors for the synthetic masks on the benchmark can be downloaded here, they have to be placed in the /eval/ directory.

Model Training

1. Download pretrained ElasitcFace model ElasticFace-Arc Model and copy it to the output folder
2. Download MS1MV2 dataset from insightface on strictly follow the licence distribution
3. Set the config.rec in config/configKD.py to the dataset path
4. Intall the requirement from requirement.txt: pip install -r requirements.txt
5. run train_kd.py

Pretrained Models

All evaluated pre-trained models are available:

• ArcFace Model
• MagFace Model
• ElasticFace-Arc Model

Our models can be downloaded here.

Citing

If you use any of the code provided in this repository or the models provided, please cite the following paper:

@INPROCEEDINGS{huber2021maskinvariant,  
   author={Huber, Marco and Boutros, Fadi and Kirchbuchner, Florian and Damer, Naser},  
   booktitle={2021 16th IEEE International Conference on Automatic Face and Gesture Recognition (FG 2021)},   
   title={Mask-invariant Face Recognition through Template-level Knowledge Distillation},   
   year={2021},  
   volume={},  
   number={},  
   pages={1-8},  
   doi={10.1109/FG52635.2021.9667081}
}

Acknowledgement

This research work has been funded by the German Federal Ministry of Education and Research and the Hessian Ministry of Higher Education, Research, Science and the Arts within their joint support of the National Research Center for Applied Cybersecurity ATHENE.

License

This project is licensed under the terms of the Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) license. Copyright (c) 2021 Fraunhofer Institute for Computer Graphics Research IGD Darmstadt