Generalized Category Discovery (GCD) aims to classify un-labeled data by leveraging knowledge from labeled cate-gories. While existing methods have achieved remarkableprogress, they often treat images as flat feature sets, neglecting the intrinsic hierarchy: where key objects dominate meaning and backgrounds serve as context. For instance, in im-ages of a dog either standing on grass or lying on a bed, thedog remains the central semantic element, whereas the back-ground varies. Motivated by this, we propose LEArning In-trinsic Hierarchy (LEAH), a lightweight module designed tomodel hierarchical structure within images. LEAH consistsof two components: a pruner that filters task irrelevant to-kens to extract key objects, and a constructor that embeds keyobjects and full images into hyperbolic space using adaptiveentailment cones to capture compositional semantics. LEAHcan be easily integrated into existing GCD frameworks withminimal modification. When applied to SimGCD, it achievesup to 13.2% accuracy improvement on fine-grained bench-marks, demonstrating its effectiveness in discovering subtleinter-class differences through hierarchical modeling.
Set paths to datasets and desired log directories in config.py
We use fine-grained benchmarks in this paper, including:
- CUB-200-2011
- Stanford Cars (Img, Annotation)
We use oarse-grained benchmarks in this paper, including:
Train the model:
bash scripts/run_${DATASET_NAME}.sh
Our results on whole five datasets:
The code repo is largely built on this repo.
If you use our code, please cite our work. :)
@inproceedings{duan2026learning,
title={Learning Intrinsic Hierarchy for Generalized Category Discovery},
author={Duan, Yu and He, Junzhi and Hu, Zhanxuan and Ji, Mengda and Wang, Rong and Gao, Quanxue},
booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
volume={40},
number={25},
pages={20950--20958},
year={2026}
}