AsArcFace: Asymmetric Additive Angular Margin Loss for Fairface Recognition
Abstract
Fairface recognition aims to the mitigate the bias between different attributes in face recognition task while maintaining the state-of-art accurancy. It is a challenging task due to high variances between different attributes and unbalancement of data. In this work, we provide an approach to make a fairface recognition by using asymmetric-arc-loss training and multi-step finetuning. First, we train a general model with an asymmetric-arc-loss, and then, we make a mutli-step finetuning to get higher auc and lower bias. Besides, we propose another viewpoint on reducing the bias and use bag of tricks such as reranking, boundary cut and hard-sample model ensembling to improve the performance. Our approach achieved the first place at ECCV 2020 ChaLearn Looking at People Fair Face Recognition Challenge.
Cite
Text
Zhou et al. "AsArcFace: Asymmetric Additive Angular Margin Loss for Fairface Recognition." European Conference on Computer Vision Workshops, 2020. doi:10.1007/978-3-030-65414-6_33Markdown
[Zhou et al. "AsArcFace: Asymmetric Additive Angular Margin Loss for Fairface Recognition." European Conference on Computer Vision Workshops, 2020.](https://mlanthology.org/eccvw/2020/zhou2020eccvw-asarcface/) doi:10.1007/978-3-030-65414-6_33BibTeX
@inproceedings{zhou2020eccvw-asarcface,
title = {{AsArcFace: Asymmetric Additive Angular Margin Loss for Fairface Recognition}},
author = {Zhou, Shengyao and Luo, Junfan and Zhou, Junkun and Ji, Xiang},
booktitle = {European Conference on Computer Vision Workshops},
year = {2020},
pages = {482-491},
doi = {10.1007/978-3-030-65414-6_33},
url = {https://mlanthology.org/eccvw/2020/zhou2020eccvw-asarcface/}
}