Augmenting Features via Contrastive Learning-Based Generative Model for Long-Tailed Classification
Thanks to the advances in deep learning-based computer vision, image classification has shown great achievements. However, it has faced a heavy class imbalance issue which is one of the characteristics of real-world datasets. The severe class imbalance makes the classifier easily biased toward majority classes and overfitting to minority classes. To address this issue, supplementing minority classes with artificially generated samples has proven effective. In addition, contrastive learning has been introduced to improve image classification performance recently. Motivated by recent works, we propose feature augmentation via a contrastive learning-based generative model for long-tailed classification. Specifically, features are augmented using the feature dictionary obtained by real samples and the generated convex weights, which are used for learning an image classification model. Here, the model for the feature augmentation is trained based on generative adversarial learning and contrastive learning in an end-to-end manner. The generative adversarial learning helps to generate real-like features, and the contrastive learning improves the feature's discrimination power. Through extensive experiments with various long-tailed classification datasets, we verify the effectiveness of the proposed method.