Class-aware Universum Inspired re-balance learning for long-tailed recognition

Abstract

Data augmentation for minority classes serves as an effective strategy for long-tailed recognition, prompting the emergence of numerous methods. Although these methods achieve balance in sample quantity, the quality of the augmented samples cannot be guaranteed, invoking issues like over-fitting, lack of variety, and semantic drift. To this end, we propose the Class-aware Universum Inspired Re-balance Learning (CaUIRL) for long-tailed recognition, which endows the Universum with class-aware ability to re-balance individual minority classes in terms of both sample quantity and quality. In particular, we theoretically prove that the classifiers learned by CaUIRL are consistent with those learned under the balanced condition from a Bayesian perspective. In addition, we develop a higher-order mixup approach, which can automatically generate class-aware Universum (CaU) data without resorting to any external data. Unlike the traditional Universum, CaU additionally takes into account domain similarity, class separability, and sample diversity into account. Comprehensive experiments on benchmark datasets reveal that the proposed method substantially enhances model performance, especially in minority classes (e.g., the top-1 accuracy of the last two tail classes is improved by 6% on Cifar10-LR).