E²IGB: Enhanced effective-information-guided class-balanced loss for long-tailed object recognition

Long-tailed object recognition is a fundamental and long-standing problem in computer vision, which suffers from performance degradation on rare categories due to severe data imbalance. Current leading approaches solve it by adjusting class loss weights based on the number of samples, however, they often overlook the redundancy in sample information, causing a mismatch between sample quantity and effective information. In this work, we propose a novel metric named Enhanced Effective Number (EEN) that more accurately reflect the effective information content of each category. Based on EEN, we further developed an Enhanced Effective-Information-Guided Balancing (E²IGB) framework for long-tailed object recognition. Specifically, E²IGB consists of two key components: an Enhanced Class-Balanced (ECB) loss that adapts loss weights according to class informativeness, and a Spearman Correlation Loss (SCL) that suppresses the undesirable positive correlation between class IoU and EEN. Extensive experiments on CIFAR-LT, ImageNet-LT, LVIS v1.0 and COCO-LT datasets are conducted and demonstrate that E²IGB consistently improves recognition performance for rare categories, while maintaining or even enhancing performance for frequent ones. Concretely, our approach achieves the state-of-the-art long-tailed object recognition performance on commonly used metrics.