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 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 of head classes, dual-classifier framework with uniform sampler performing duties. For balancing learning feature representation classifiers, dynamic weight is introduced to adjust model’s attention. alleviate deviation between training data testing data, hierarchical rebalancing loss designed re-weighting branch, adjusts decision values predicted logits facilitate actively compensating categories. Finally, conduct extensive experiments standard benchmarks Cifar10-LT, Cifar100-LT, ImageNet-LT, iNaturalist2018, demonstrating effectiveness superiority our HRDC.