Long-Tailed Recognition by Hierarchical Rebalancing Dual-Classifier

Image classification techniques have succeeded greatly on various large-scale visual datasets using deep convolution neural networks. However, previous deep models usually suffer severe performance degradation in highly skewed datasets, which restricts their practical application. In this paper, we propose a novel Hierarchical Rebalancing Dual-Classifier model for long-tailed recognition. To better identify the tail samples and maintain the performance of head classes, we propose a dual-classifier framework with a uniform sampler for performing their duties. For balancing the learning of feature representation and classifiers, a dynamic weight is introduced to adjust the model's attention. To alleviate the feature deviation between training data and testing data, a hierarchical rebalancing loss is designed for the re-weighting branch, which adjusts the decision values in predicted logits to facilitate the model actively compensating for tail categories. Finally, we conduct extensive experiments on standard long-tailed benchmarks Cifar10-LT, Cifar100-LT, ImageNet-LT, and iNaturalist2018, demonstrating the effectiveness and superiority of our HRDC.