Location-Based Scene Reconstruction for Long-Tail Recognition

Real-world data often exhibit a long-tailed distribution with severe class imbalance. In such cases, the majority class dominates the deep learning training, which changes the decision boundary of the minority class and reduces the classification accuracy. In this paper, we propose a novel location-based scene reconstruction(LSR) data augmentation method for long-tail recognition. This approach uses a gradient localization method to increase the scenes of tail class samples and enhance the discrimination of the model between head and tail classes, thus the accuracy of long-tail recognition is improved. Experiments on two benchmark datasets show that the LSR method achieves state-of-the-art performance on the long-tail recognition task. More importantly, our method can be easily combined with other classification methods and improves the performance of these traditional classification methods