Generalized feature augmentation：feature space object recognition algorithm for generalized long-tail classification

In practical scenarios, the frequencies of each class in a dataset typically differ, often following a long-tailed distribution. To alleviate data imbalance issues, methods such as resampling, re-weighting, and data augmentation are commonly employed to ensure relatively consistent performance across classes during model training. When samples for certain classes in the dataset are insufficient to adequately represent that class, additional knowledge is necessary to recover lost information. Most existing long-tail classification methods focus primarily on addressing imbalances in class sample sizes, while overlooking imbalances in attributes. However, even in scenarios of class balance, attribute distributions may exhibit long-tailed patterns, and generalized long-tail classification methods address both forms of imbalance simultaneously. This study proposed a novel approach to address the issue of inadequate sample representation in generalized long-tail distributions by augmenting features for underrepresented classes with shared and class-specific features learned from classes with sufficient samples. Specifically, invariant features unrelated to attributes were extracted and decomposed into class-shared and class-specific components using class activation maps. During training, the class-specific features for underrepresented classes were fused with class-shared features to dynamically generate new samples for those classes. Experimental results on generalized long-tailed datasets such as ImageNet-GLT and MSCOCO-GLT demonstrate advanced performance.