How Does Data Augmentation Affect to Model Performance in Long-Tailed Classification?

Abstract

Long-tailed classification is one of the biggest issues in the real-world, because severe data imbalances often lead to less accurate forecasts in the minority. This seriously affects deploying prediction models on recognition systems and embedded devices. Most recently methods have focused on improving the model’s performance by proposing new rebalance strategies, using more networks, transfer learning, etc. In this paper, we investigate the impact of data augmentation for long-tailed classification via two proposed cases including data augmentation for all classes and data augmentation only for minority classes. Data augmentation intuitively increases the imbalance ratio between majority classes and minority classes. This may directly affect the model’s performance. However, the experiment results in case 1 have shown the opposite, we found that with more data at tailed classes, the model can learn better and achieve higher despite the imbalance ratio increase. In addition, it serves as an updated version of case 1 to overcome its limitations. The results have depicted that our approach achieves state-of-the-art performance on long-tailed classification.