Prior-constrained Association Learning for Fine-grained Generalized Category Discovery

February 26, 2025 ยท View on GitHub

This repo contains the implementation code of our paper: Prior-constrained Association Learning for Fine-grained Generalized Category Discovery.

teaser The proposed method PAL-GCD is primarily focused on non-parametric classification through prototypical contrastive learning and prior-constrained data association. Additionally, it also provides the combination of parametric and non-parametric classification by which a higher performance can be obtained.

Running

Dependencies

pip install -r requirements.txt

Config

Set paths to datasets and desired log directories in config.py

Datasets

We use fine-grained benchmarks in this paper, including:

We also use generic object recognition datasets, including:

Scripts

Train the model:

  • Train with only non-parametric classifier:
sh run_${DATASET_NAME}_stage1.sh
  • Train with joint parametric and non-parametric classifier:
sh run_${DATASET_NAME}_stage1_and_stage2.sh

Citation

If you find this repo useful for your research, please consider citing our paper:

@inproceedings{wang2025palGCD,
  title={Prior-constrained Association Learning for Fine-grained Generalized Category Discovery},
  author={Menglin Wang and Zhun Zhong and Xiaojin Gong},
  booktitle={AAAI Conference on Artificial Intelligence},
  year={2025}
}

Acknowledgements

The codebase is largely built on this repo: https://github.com/CVMI-Lab/SimGCD. Thanks to the authors for their method implementation.

License

This project is licensed under the MIT License - see the LICENSE file for details.