DyFo: Test Code and Datasets

This repository contains the test code and evaluation data for DyFo: A Training-Free Dynamic Focus Visual Search for Enhancing LMMs in Fine-Grained Visual Understanding (CVPR 2025 Highlight).

For the full implementation, please refer to the main codebase at:
👉 https://github.com/PKU-ICST-MIPL/DyFo_CVPR2025

📦 How to Use

1. Download and unzip this dataset into the root directory of the main repository.
2. This package includes evaluation code and processed data for testing the DyFo method.

📚 Included Datasets

The following benchmark datasets are included for evaluation:

• POPE-COCO/A-OKVQA/GQA
  Original source: https://github.com/RUCAIBox/POPE

• Vstar
  Original source: https://github.com/penghao-wu/vstar

These datasets are pre-processed and ready to use with our test code.

📄 License

This dataset and code are licensed under the Apache 2.0 License.

📑 Citation

If you use DyFo in your work, please cite:

@misc{li2025dyfotrainingfreedynamicfocus,
  title={DyFo: A Training-Free Dynamic Focus Visual Search for Enhancing LMMs in Fine-Grained Visual Understanding},
  author={Geng Li and Jinglin Xu and Yunzhen Zhao and Yuxin Peng},
  year={2025},
  eprint={2504.14920},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2504.14920}
}