Improving Representation Consistency with Pairwise Loss for Masked Face Recognition

Hanjie Qian, Panpan Zhang, Sijie Ji, Shuxin Cao, Yuecong Xu

Abstract: Given the coronavirus disease (COVID-19) pandemic, people need to wear masks to protect themselves and reduce the spread of COVID, which brings new challenge to the traditional face recognition task. Since features like the nose and mouth, which are well distinguishable, are hidden under the mask, traditional methods are no longer simply applicable, even though they once achieved a high degree of accuracy. In response to this problem, the Masked Face Recognition Challenge & Workshop (MFR) was held in conjunction with the International Conference on Computer Vision (ICCV) 2021. This article details a method that combining the classic ArcFace and pairwise loss to target the new masked face recognition task. So far, our method has achieved the second place in the competition.