Nov 2026· Proceedings of the 35th ACM International Conference on Information and Knowledge Management· 0 citations· 47 references
TL;DR
This paper proposes VISRA, a semantic-driven framework for LT-GCD built upon a parametric GCD pipeline that leverages Vision-Language Models to provide additional semantic supervision for representative novel-class instances and guide visual representation learning.
Abstract
Generalized Category Discovery (GCD) aims to classify unlabeled data containing both known and novel classes, given only labels for known classes. Existing GCD methods typically assume a relatively balanced data distribution, an assumption that breaks down in realistic open-world scenarios. In this paper, we study Long-Tailed Generalized Category Discovery (LT-GCD), where both labeled and unlabeled data follow long-tailed distributions. Under such conditions, training signals are dominated by known categories; this imbalance biases representation learning and weakens novel-category discovery. To address these challenges, we propose VISRA, a semantic-driven framework for LT-GCD built upon a parametric GCD pipeline. Rather than relying solely on distribution debiasing, our framework leverages Vision-Language Models (VLMs) to provide additional semantic supervision for representative novel-class instances and guide visual representation learning. Experimental results on three long-tailed datasets demonstrate that VISRA achieves state-of-the-art performance, effectively alleviating novel-class supervision deficiency and learning more discriminative representations under severe imbalance. Our code and prompt are available at https://github.com/EileanZYZ/VISRA.
Adapting vision-language models to downstream tasks has achieved remarkable success by leveraging pseudo-labels generated from unlabeled data. Existing methods typically assume a uniform unlabeled data distribution, and thus the resulting pseudo-label distribution is likewise uniform. However, real-world data distribut...
Ke-Liang Chen, Ya-Xin Hou, Hui Liu et al.· 0 citations
This work proposes Domain-Context Prompting (DCP), a zero-shot framework that calibrates the full set of ID textual prototypes with shared domain semantics and demonstrates the effectiveness of DCP in the zero-shot setting and further improvements from its few-shot extension.
Qi-Chao Chen, Yan-Qi Wu, Zhi-Yuan Chen et al.· IEEE Signal Processing Lette...· 0 citations
Test-Time Generalized Category Discovery (TT-GCD), a unified and more realistic scenario where a vision-language model must adapt to distribution shifts, classify known categories using only textual supervision, and discover novel categories, all during test time and without access to labeled data is introduced.
Shambhavi Mishra, Omprakash Chakraborty, Julio Silva-Rodríguez et al.· arXiv.org· 0 citations
PAIR (Prototype-Aware Inertia Rectification) is proposed, a unified framework that reduces seen-category bias from both local and global perspectives and provides a more balanced basis for open-vocabulary prediction.
Jin-Yi Fang, Bing-Ke Zhu, Jing-Ling Yuan et al.· Proceedings of the 34th ACM...· 0 citations
Learning with long-tailed noisy labels presents a fundamental challenge for classification, as the combined effects of class imbalance and label corruption distort the feature space. These distortions lead to two forms of inseparability: i) semantic confusion, where mislabeled samples resemble semantically similar clas...
Jie Wang, Yi-Jie Lin, Ziang Niu et al.· Proceedings of the 34th ACM...· 0 citations
A compact explainability-driven framework for comparing encoders trained with classification-only and caption-generation supervision on CUB-200 provides a systematic basis for assessing learned representations and comparing alternative training and transfer-learning strategies.
Md Raisul Kibria, Sébastien Lafond· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.