Sampling-tailored two-pronged network for long-tailed class imbalance learning

Abstract

A long-tail class imbalanced learning problem is a scenario where the rare or minority classes, representing infrequent events or categories, make up the long tail of the class distribution and have disproportionately few examples compared to the dominant classes. The resulting imbalance makes it challenging to train models effectively for these underrepresented classes. We introduce a comprehensive solution - STTP-Net: Sampling-Tailored Two-Pronged Network for long-tail class-imbalanced learning, which aims to address this issue holistically. The study thoroughly examines mixed sample data augmentation techniques in conjunction with various sampling strategies to identify the most effective approaches for handling long-tail imbalance. Based on this analysis, a hybrid mixup strategy tailored explicitly for data augmentation in long-tail imbalanced settings is proposed. The core of the proposed approach comprises a two-pronged network consisting of two classification heads designed to handle long-tail imbalanced datasets. One head specializes in learning the head and median classes in this design. In contrast, the other head becomes an expert in tail classes, striking a balance between accurate prediction of tail classes without compromising accuracy for the head classes. Additionally, we address the label distribution shifts in long-tail imbalance by introducing an Effective Balanced Softmax (EBS) function. The presented method achieves state-of-the-art performance in several benchmark categories for long-tail visual recognition datasets, surpassing the most prominent and pertinent end-to-end and dual-branch approaches.