Griffon v2 12M Dataset Card

News

** [2025/08/10] ** We are happy to announce that Griffon v2 is accepted to ICCV 2025.

Dataset details

We provide 12M data used in the stage 2 training of Griffon v2. In this repo, we provide the processed annotation files for Obejct Detection, REC/REG, Visual Grounding, and Non-existing Judging tasks in the paper, and also the self-collected object counting data.

Self-Counting Data

The counting data includes three parts, CT-datasets-new.tar.gz, CountAnythingV1_clean.tar.gz, and train_visual_openimages_cocostyle_cls601.json. For the file ends with tar.gz, it contains both the annotation file and also the images. While for the data collected from the OpenImages, please download the OpenImages2019 train images.

Other Data

For other annotations, please download the images from the following datasets: COCO(train2014 & train2017), Visual Genemo, Objects365-2023, V3Det, and Flickrs30K Entities. If meet any problem like missing images, please contact us from the Github issues.

License

Attribution-NonCommercial 4.0 International. It should abide by the policy of the original data sources.

Citation

To use this data, please cite

@article{zhan2024griffon,
  title={Griffon v2: Advancing multimodal perception with high-resolution scaling and visual-language co-referring},
  author={Zhan, Yufei and Zhu, Yousong and Zhao, Hongyin and Yang, Fan and Tang, Ming and Wang, Jinqiao},
  journal={arXiv preprint arXiv:2403.09333},
  year={2024}
}