A Comprehensive Study on Loss Functions for Cross-Factor Face Recognition

Gee-Sern Jison Hsu, Hung-Yi Wu, Moi Hoon Yap; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2020, pp. 826-827

Abstract


A significant progress has been made to face recognition in recent years. The progress includes the advancement of the deep learning solutions and the availability of more challenging databases. As the performance on previous benchmark databases, such as MPIE and LFW, saturates, more challenging databases are emerging and keep driving the development of face recognition technology. The loss function considered in a deep face recognition network plays a critical role for the performance. To better evaluate the state-of-the-art loss functions, we define four challenging factors, including pose, age, occlusion and resolution with specific databases and conduct an extensive experimental study on the latest loss functions. We select the IARPA Janus Benchmark-B (IJB-B) and IARPA Janus Benchmark-C (IJB-C) for pose, the FG-Net Aging Database (FG-Net) for age, the AR Face Database (AR Face) for occlusion, and the Surveillance Cameras Face Database (SCface) for low resolution. The loss functions include the Center Loss, the Marginal Loss, the SphereFace, the CosFace and the ArcFace. Although for most factors, the ArcFace outperforms others. However, the best performance against low-resolution is achieved by the SphereFace. Another attractive finding of this study is that the cross-age performance is the lowest among the four factors with a clear margin. This highlight possible directions for future research.

Related Material


[pdf]
[bibtex]
@InProceedings{Hsu_2020_CVPR_Workshops,
author = {Hsu, Gee-Sern Jison and Wu, Hung-Yi and Yap, Moi Hoon},
title = {A Comprehensive Study on Loss Functions for Cross-Factor Face Recognition},
booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops},
month = {June},
year = {2020}
}