| Summary: | With the rapid increase in data scale, real-world datasets tend to exhibit long-tailed class distributions (i.e., a few classes account for most of the data, while most classes contain only a few data points). General solutions typically exploit class rebalancing strategies involving resampling and reweighting based on the sample number for each class. In this work, we explore an orthogonal direction, category splitting, which is motivated by the empirical observation that naive splitting of majority samples could alleviate the heavy imbalance between majority and minority classes. To this end, we propose a novel classwise splitting (CWS) method built upon a dynamic cluster, where classwise prototypes are updated using a moving average technique. CWS generates intra-class pseudo labels for splitting intra-class samples based on the point-to-point distance. Moreover, a group mapping module was developed to recover the ground truth of the training samples. CWS can be plugged into any existing method as a complement. Comprehensive experiments were conducted on artificially induced long-tailed image classification datasets, such as CIFAR-10-LT, CIFAR-100-LT, and OCTMNIST. Our results show that when trained with the proposed class-balanced loss, the network is able to achieve significant performance gains on long-tailed datasets.
|
|---|