Long-tailed image recognition through balancing discriminant quality

Abstract

Long-tailed image recognition is a challenging task in real scenes with large-scale data. Popular strategies, such as loss reweighting and data resampling, aim to reduce the model bias toward head classes. Specifically, different loss reweighting approaches explore various endogenous or exogenous measures. In this paper, we study a new endogenous measure called discriminant quality (DQ) by considering validation accuracy and discriminant uncertainty. DQ takes advantage of continuous information over a period of time. It is more robust than instantaneous information because of the mitigation of measuring instability caused by random perturbations during training. Additionally, the weight of each class is automatically rebalanced based on DQ. Consequently, the class weight supports the design of a dynamic updating strategy for the significance of the DQ difference. Experiments on MNIST-LT, CIFAR-100-LT, ImageNet-LT, and Places-LT demonstrated the superiority of DQ over state-of-the-art ones in terms of prediction accuracy.