Generalized Class Discovery (GCD) is a crucial task in machine learning that aims to identify both known and unknown classes from unlabeled datasets. However, existing GCD methods often assume that the occurrence of categories in the unlabeled data is evenly distributed, which is not the case in real-world scenarios. In natural environments, visual classes typically exhibit a long-tailed distribution, with some categories being much more prevalent than others.
This article introduces the concept of Long-tailed Generalized Category Discovery (Long-tailed GCD), which addresses the limitations of prevailing GCD methods by taking into account the imbalanced nature of real-world unlabeled datasets. The authors propose a robust methodology that incorporates two strategic regularizations to tackle the unique challenges posed by long-tailed GCD.
First, the authors propose a reweighting mechanism that increases the prominence of less-represented, tail-end categories. This approach acknowledges that rare categories are often vital but overlooked in traditional GCD methods. By assigning higher weights to these underrepresented categories during training, the proposed method ensures that they receive sufficient attention and are not overshadowed by the more frequent categories.
Second, a class prior constraint is introduced, which aligns with the expected class distribution in long-tailed datasets. This constraint takes into account the knowledge that certain categories are more likely to occur than others and incorporates this information into the GCD framework.
To evaluate the effectiveness of their proposed method, the authors conducted comprehensive experiments on two benchmark datasets: ImageNet100 and CIFAR100. The results demonstrated that their method outperformed previous state-of-the-art GCD methods by achieving an improvement of approximately 6-9% on ImageNet100. Furthermore, it achieved competitive performance on CIFAR100, demonstrating its effectiveness in handling long-tailed GCD scenarios.
Overall, this research makes a significant contribution to the field of Generalized Class Discovery by addressing the limitations of existing methods in handling long-tailed datasets. The proposed methodology, with its reweighting mechanism and class prior constraint, provides a more accurate and robust approach for discovering categories in real-world unlabeled datasets. Future research could explore further enhancements to this approach, such as incorporating additional information sources or adapting it to other domains beyond computer vision.