Novel Class Discovery for Long-tailed Recognition

While the novel class discovery has achieved great success, existing methods usually evaluate their algorithms on balanced datasets. However, in real-world visual recognition tasks, the class distribution of a dataset is often long-tailed, making it challenging to apply those methods. In this paper, we propose a more realistic setting for novel class discovery where the distribution of novel and known classes is long-tailed. The challenge of this new problem is to discover novel classes with the help of known classes under an imbalanced class scenario. To discover imbalanced novel classes efficiently, we propose an adaptive self-labeling strategy based on an equiangular prototype representation. Our method infers better pseudo-labels for the novel classes by solving a relaxed optimal transport problem and effectively mitigates the biases in learning the known and novel classes. The extensive results on CIFAR100, ImageNet100, and the challenging Herbarium19 and large-scale iNaturalist18 datasets demonstrate the superiority of our method.