Learning to resolve uncertainties for large-scale face recognition

Abstract

• We propose a ranking loss regularization for large-scale face recognition by reducing uncertainties during the training phase. • This work studies the effect of relabeling mechanism by integrating ranking loss with the ArcFace loss function. • This work demonstrates the effectiveness of the proposed method through experiments on public face recognition benchmarks. Facial recognition is a category of biometric security, used widely in various industries where we identify and authenticate an individuals identity using their face. In the modern deep learning era, face recognition datasets are playing a significant role in achieving state-of-the-art accuracy by acquiring and training millions of face images. Annotating such a large-scale face recognition dataset is challenging due to low-quality face images, and incorrect annotations unknowingly made by annotators . Training a deep learning model with such uncertainties leads to deep model overfitting on noisy uncertain samples and degradation of the discriminative ability of the model. To address these issues, we propose a simple yet effective uncertainty learning network that efficiently reduces over-fitting caused by uncertain face images. More specifically, our FC module weights each sample in the mini-batch at the decision layer, and relabeling mechanism carefully modify the labels of incorrect samples in the mini- batch. Results on IJB-B, IJB-C, LFW, AgeDB30, CFP-FP, CALFW and CPLFW public datasets demonstrate that our approach achieves state-of-the-art performance