Contrastive Learning with Hallucinating Data for Long-Tailed Face Recognition

Face recognition has been well studied over the past decades. Most existing methods focus on optimizing the loss functions or improving the feature embedding networks. However, the long-tailed distribution problem, i.e, most of the samples belong to a few identities, while the remaining identities only have limited samples, is less explored, where these datasets are not fully utilized. In this paper, we propose a learning framework to balance the long-tailed distribution problem in public face datasets. The proposed framework learns the diversity from head identity samples to generate more samples for identifying persons’ identities in the tail. The generated samples are used to finetune face recognition models through a contrastive learning process. The proposed framework can be adapted to any feature embedding networks or combined with different loss functions. Experiments on both constrained and unconstrained datasets have proved the efficiency of the proposed framework.