haoosz / FFR-Net

PyTorch code for ICIP 2022 paper "A Unified Framework for Masked and Mask-Free Face Recognition via Feature Rectification"

FFR-Net

The work of "A Unified Framework for Masked and Mask-Free Face Recognition via Feature Rectification".

Introduction

We introduce a unified framework, named Face Feature Rectification Network (FFR-Net), for masked and mask- free face recognition. Experiments show that our frame- work achieves the best average performance on mixed datasets (mask and mask-free faces) with a single model. Besides, we propose rectification blocks (RecBlocks) to rectify features of masked or mask-free faces in both spatial and channel dimensions. RecBlocks can maximize the consistency between masked faces and their mask-free counterparts in the rectified feature space.

Environment

The work is with Python 3.6 and PyTorch 1.7.

Pretrained Models

We provide 2 pretrained models:

models	pretrained file	mask-free	masked
SENet50	se50.pth	✔	✘
FFR-Net	FFRNet.pth	✔	✔

Our model is trained on augmented CASIA-WebFace dataset. Note that the pretrained weight of SENet is fixed during the training process.

Data

We provide masked CASIA-WebFace and masked LFW datasets (already aligned) and corresponding txt files:

CASIA-WebFace dataset	CASIA-WebFace clean list	LFW dataset	LFW pair list
casia_masked.tar.gz	casia_cleanlist.txt	lfw_masked.tar.gz	lfw_pairs.txt

Quick start

Type the following commands to train the model:

python3 run.py

License

This work is under MIT License.

Citation

If you find this code useful in your research, please consider citing:

@inproceedings{hao2022unified,
  title={A Unified Framework for Masked and Mask-Free Face Recognition via Feature Rectification},
  author={Hao, Shaozhe and Chen, Chaofeng and Chen, Zhenfang and Wong, Kwan-Yee K},
  booktitle={2022 IEEE International Conference on Image Processing (ICIP)},
  year={2022}
}