LONG-TAILED RECOGNITION BY LEARNING FROM LATENT CATEGORIES
In this work, we address the challenging task of long-tailed recognition. Previous long-tailed recognition methods commonly focus on data augmentation of tailed classes or re-balancing strategy to give more attention to tailed classes during training. However, due to the limited training images for tailed classes, the diversity of augmented images are still restricted, which results in poor feature representations. In this work, we argue that there are common latent features between the head and tailed classes that can be used to give better feature representation. We propose to learn a set of semantic and class-agnostic latent features shared by the head and tailed classes. Then, we implicitly enrich the training sample diversity via leveraging semantic data augmentation for the commonality features. We evaluate our methods on several popular long-tailed datasets and achieve new state-of-the-art performance consistently.