MaxSwap-Enhanced Knowledge Consistency Learning for long-tailed recognition

Abstract

Deep learning has made significant progress in image classification. However, real-world datasets often exhibit a long-tailed distribution, where a few head classes dominate while many tail classes have very few samples. This imbalance leads to poor performance on tail classes. To address this issue, we propose MaxSwap-Enhanced Knowledge Consistency Learning which includes two core components: Knowledge Consistency Learning and MaxSwap for Confusion Suppression. Knowledge Consistency Learning leverages the outputs from different augmented views as soft labels to capture inter-class similarities and introduces a consistency constraint to enforce output consistency across different perturbations, which enables tail classes to effectively learn from head classes with similar features. To alleviate the bias towards head classes, we further propose a MaxSwap for Confusion Suppression to adaptively adjust the soft labels when the model makes incorrect predictions which mitigates overconfidence in incorrect predictions. Experimental results demonstrate that our method achieves significant improvements on long-tailed datasets such as CIFAR10-LT, CIFAR100-LT, ImageNet-LT, and Places-LT, which validates the effectiveness of our approach.