WeightFace: weight adaptive scaling loss for face recognition

The discriminability between distinct classes in the embedding space is improved by margin-based loss functions like SphereFace, CosFace and Arcface. More recently, face quality is introduced to face recognition to adaptively adjust the margin. However, these methods ignore the class imbalance problem, which affects the distribution of each class in the real embedding space and misleads the classification results of minority class. In this paper, a novel loss (WeightFace) is proposed to learn the scale parameter adaptively guided by class weight to address the class imbalance problem. We have proved that the minority class requires a larger scale parameter, since the minority class get a relatively smaller gradient so that the minority class has an intra-class variation comparable to the majority class. The weight magnitude is used to connect classes and scale parameters. This allows different classes distributing properly in real space and the test accuracy is boosted. Extensive experiments on popular benchmarks demonstrate the superiority of our WeightFace over state-of-the-arts.