# Dataset Summary

ADOPD is a large-scale dataset for document page decomposition, distinguished by a novel data-driven document taxonomy discovery method for data collection. This approach combines large-scale pretrained models with a human-in-the-loop process to ensure diversity and balance in the data. ADOPD includes densely annotated labels for document images, covering four tasks: Doc2Mask, Doc2Box, Doc2Tag, and Doc2Seq. Annotations for each image include human-labeled entity masks, text bounding boxes, and automatically generated tags and captions. Detailed experimental analyses validate the data-driven document taxonomy method and evaluate the four tasks using different models. ADOPD aims to support future research in document image understanding.

![](top_ADOPD_Tasks.png)

# Dataset Information
The ADOPD dataset contains a total of 120,000 images, with the following language distribution:
- English: 60,000
- Chinese: 20,000
- Japanese: 20,000
- Others: 20,000

# Citation
```
@inproceedings{
    gu2024adopd,
    title={{AD}o{PD}: A Large-Scale Document Page Decomposition Dataset},
    author={Jiuxiang Gu and Xiangxi Shi and Jason Kuen and Lu Qi and Ruiyi Zhang and Anqi Liu and Ani Nenkova and Tong Sun},
    booktitle={The Twelfth International Conference on Learning Representations},
    year={2024},
    url={https://openreview.net/forum?id=x1ptaXpOYa}
}
```

# License (cc-by-nc-nd-4.0)