Attentive Feature Augmentation for Long-Tailed Visual Recognition

Deep neural networks have achieved great success on many visual recognition tasks. However, training data with a long-tailed distribution dramatically degenerates the performance of models. In order to relieve this imbalance problem, an effective Long-Tailed Visual Recognition (LTVR) framework is proposed based learned balance and robust features under circumstances. framework, plug-and-play Attentive Feature Augmentation (AFA) module designed mine class-related variation-related original samples via novel hierarchical channel attention mechanism. Then, those are aggregated synthesize fake cope dataset. Moreover, Lay-Back Learning Schedule (LBLS) developed ensure good initialization feature embedding. Extensive experiments conducted two-stage method verify effectiveness both learning classifier rebalancing in image task. Experimental results show that, when trained imbalanced datasets, achieves superior over state-of-the-art methods.