Unknown Identity Rejection Loss: Utilizing Unlabeled Data for Face Recognition

Abstract

Face recognition has advanced considerably with the availability of large-scale labeled datasets. However, how to further improve the performance with the easily accessible unlabeled dataset remains a challenge. In this paper, we propose the novel Unknown Identity Rejection (UIR) loss to utilize the unlabeled data. We categorize identities in unconstrained environment into the known set and the unknown set. The former corresponds to the identities that appear in the labeled training dataset while the latter is its complementary set. Besides training the model to accurately classify the known identities, we also force the model to reject unknown identities provided by the unlabeled dataset via our proposed UIR loss. In order to 'reject' faces of unknown identities, centers of the known identities are forced to keep enough margin from centers of unknown identities which are assumed to be approximated by the features of their samples. By this means, the discriminativeness of the face representations can be enhanced. Experimental results demonstrate that our approach can provide obvious performance improvement by utilizing the unlabeled data.

Cartoon Face Recognition

“…The UIR loss L aims to find a feature re-projection with unsupervised regularization between different domains. Inspired by [33], we develop the unknown identity rejection loss L , for each batch sample B , taking 1/4 human face data x ℎ as unlabeled data and 3/4 of x as known ones. L ℎ is defined respectively.…”

Asymmetric Rejection Loss for Fairer Face Recognition

“…Unknown Identity Rejection Loss [29] tries to minimize the negative entropy for the classification probability distribution of unlabeled images. Their loss is rewritten in Equation 1.…”

“…If clusters contain noise, as pointed by [21], models' performance could be dampened a lot. Therefore, unlike Transductive Centroid Projection (TCP) [14], we don't take clustering into consideration and we leverage the unlabeled data in a similar way as [29]. In [29], the authors design an Unknown Identity Rejection (UIR) loss that requires unlabeled data being "rejected" by annotated classes in labeled training dataset, shown in Figure 1(b).…”

“…Therefore, unlike Transductive Centroid Projection (TCP) [14], we don't take clustering into consideration and we leverage the unlabeled data in a similar way as [29]. In [29], the authors design an Unknown Identity Rejection (UIR) loss that requires unlabeled data being "rejected" by annotated classes in labeled training dataset, shown in Figure 1(b). In this paper, we point out the weakness of their method, and modify it for the first step to alleviate aforementioned model racial bias.…”

“…In this study, following TCP [14] and UIR [29], we firstly filter out the possible overlap between labeled dataset and unlabeled images. By viewing each unlabeled image as a unique class, unlabeled data are utilized in a supervised way as [14], but in our approach unlabeled data and labeled data are treated asymmetrically.…”