Generalized Category Discovery with Decoupled Prototypical Network

Generalized Category Discovery (GCD) aims to recognize both known and novel categories from a set of unlabeled data, based on another dataset labeled with only categories. Without considering differences between categories, current methods learn about them in coupled manner, which can hurt model's generalization discriminative ability. Furthermore, the training approach prevents these models transferring category-specific knowledge explicitly data lose high-level semantic information impair model performance. To mitigate above limitations, we present called Decoupled Prototypical Network (DPN). By formulating bipartite matching problem for category prototypes, DPN not decouple achieve different targets effectively, but also align transfer capture semantics. more features through our proposed Semantic-aware Learning (SPL). Besides capturing meaningful information, SPL alleviate noise hard pseudo labels semantic-weighted soft assignment. Extensive experiments show that outperforms state-of-the-art by large margin all evaluation metrics across multiple benchmark datasets. Code are available at https://github.com/Lackel/DPN.