Class Activation Maps-based Feature Augmentation for long-tailed classification

Abstract

It remains an important and challenging problem for the classification of long-tailed data. Most existing methods focus on sampling strategies based on tail classes. However, adequately representing tail classes becomes challenging when they contain only a few sparse samples, leading to difficulty in capturing accurate decision boundaries. To deal with this problem, we propose a Layer Class Activation Map for feature augmentation (LayerCAM-FA) method. The core of our method is to make full use of the knowledge of head classes to augment the features of tail classes. First, the layer-class activation map is employed to capture the features learned from a network which is most useful for identifying the label. The extracted features are decomposed into two parts: distinctive features and common features. The distinctive features pay more attention to recognizing the label, and the common features concentrate on generalized information of different classes. Then, some head classes with similar characteristics to tail classes are selected by attribute and probability distance to augment the features of tail classes. Finally, the distinctive features of tail classes are combined with the common features of head classes to augment the features of tail classes, and the augmented features are used to fine-tune the network to obtain the final classification result. The effectiveness of our proposed LayerCAM-FA is demonstrated using experimental results compared with other related methods.