We release part of our experimental code as a reference for readers that are interested in our paper: Unknown Identity Rejection Loss: Utilizing Unlabeled Data for Face Recognition
We proposed the Unseen Identity Rejection Loss to utilize the unlabeled data for training discriminative feature representations for face recognition. By introducing features of unknown faces into the high-dimensional space, the feature space can be rearranged so that margins between samples can be further optimized.
Below is the performance of our trained ResNet-100 model on the IJB-C dataset.
The implementation of the loss function is in lines 252~265 of selelcted_code/semisupervised-finetune.py
. Our implementation is based on the InsightFace repository. Here we used ResNet-100 as the backbone network and arcface loss function for supervised training.
If you find our method useful in your research, please consider citing the following related paper.
@article{uirloss,
title={Unknown Identity Rejection Loss: Utilizing Unlabeled Data for Face Recognition},
author={H. Yu and Y. Fan and K. Chen and Y. He and X. Lu and J. Liu and D. Xie},
journal={ICCV Workshop},
year={2019}
}