# Wild-OmniDocBench

**A Real-World Captured Document Parsing Benchmark for Robustness Evaluation**

<p align="center">
  <a href="https://huggingface.co/datasets/VirtualLUO/Wild_OmniDocBench/blob/main/README_ZH.md">中文版</a> •
  <a href="https://arxiv.org/abs/2603.23885">Paper</a> •
  <a href="https://github.com/VirtualLUOUCAS/Wild_OmniDocBench">GitHub</a> •
  <a href="https://huggingface.co/datasets/VirtualLUO/Wild_OmniDocBench">HuggingFace</a>
</p>

## Overview

**Wild-OmniDocBench** is a benchmark for evaluating document parsing robustness under real-world captured conditions. It is derived from [OmniDocBench](https://github.com/opendatalab/OmniDocBench) by converting scanned/digital documents into naturally captured images through controlled physical simulation, including printing, deformation, and photography under diverse lighting conditions.

Unlike standard benchmarks that rely on clean scanned or digital-born pages, Wild-OmniDocBench introduces realistic artifacts such as:
- **Geometric distortions** (perspective shifts, bends, wrinkles)
- **Illumination variations** (directional, uneven, low-light)
- **Screen capture artifacts** (moire patterns, reflections)
- **Environmental interference** (background overlays, shadows)

> **Note:** The current release of Wild-OmniDocBench corresponds to **OmniDocBench v1.5**. We are currently processing the extended portions for v1.6 and will release them in a future update.

<p align="center">
  <img src="assets/overview.png" width="90%" alt="Wild-OmniDocBench Construction">
</p>

## Benchmark Statistics

| Item | Details |
|------|---------|
| Total Images | 1,350 |
| Source | Real-world captured variant of OmniDocBench |
| Document Types | Books, Textbooks, Papers, PPTs, Newspapers, Notes, Exams, Magazines, Financial Reports, etc. |
| Capture Methods | (i) Print + physical deformation + photography; (ii) Screen display + re-capture |
| Annotations | Inherited from OmniDocBench (full structural and reading-order annotations) |

## Data Format

### Directory Structure

```
Wild_OmniDocBench/
├── README.md                   # English README
├── README_ZH.md                # Chinese README
├── wild_omnidocbench.zip       # Benchmark images (1,350 JPGs)
└── assets/
    └── overview.png            # Overview figure
```

### Images

After unzipping `wild_omnidocbench.zip`, images are named following the OmniDocBench convention:

```
{doc_type}_{language}_{source}_{page}.jpg
```

For example: `book_en_A.Concise.Introduction.to.Linear.Algebra_page_065.jpg`

## Evaluation

Wild-OmniDocBench uses the same annotation format and evaluation protocol as [OmniDocBench](https://github.com/opendatalab/OmniDocBench). To evaluate on Wild-OmniDocBench:

1. **Obtain annotations and evaluation scripts** from the official OmniDocBench repository:
   ```
   https://github.com/opendatalab/OmniDocBench
   ```

2. **Replace the image source** with Wild-OmniDocBench images (from `wild_omnidocbench.zip`).

3. **Run evaluation** following the OmniDocBench protocol. Metrics include:
   - **Overall Score** (↑)
   - **Text Edit Distance** (↓)
   - **Formula CDM** (↑)
   - **Table TEDS** (↑)
   - **Reading Order Edit Distance** (↓)

## Key Results

Performance degradation from OmniDocBench to Wild-OmniDocBench (from the DocHumming paper):

| Model | Type | Overall (Origin) | Overall (Wild) | Degradation |
|-------|------|:-:|:-:|:-:|
| DocHumming (1B) | End2End | 93.75 | 87.03 | −6.72 |
| dots.ocr (3B) | End2End | 88.41 | 78.01 | −10.40 |
| Qwen3-VL (235B) | General | 89.15 | 79.69 | −9.46 |
| MinerU2.5 (1.2B) | Modular | 90.67 | 70.91 | −19.76 |
| PaddleOCR-VL (0.9B) | Modular | 91.93 | 72.19 | −19.74 |

End-to-end models exhibit significantly less degradation than modular cascaded pipelines under real-world capture conditions.

## Citation

```bibtex
@misc{li2026towardsrealworlddocument,
      title={Towards Real-World Document Parsing via Realistic Scene Synthesis and Document-Aware Training},
      author={Gengluo Li and Pengyuan Lyu and Chengquan Zhang and Huawen Shen and Liang Wu and Xingyu Wan and Gangyan Zeng and Han Hu and Can Ma and Yu Zhou},
      year={2026},
      journal={arXiv preprint arXiv:2603.23885},
      url={https://arxiv.org/abs/2603.23885},
}
```

## Acknowledgements

Wild-OmniDocBench is built upon [OmniDocBench](https://github.com/opendatalab/OmniDocBench). We thank the OmniDocBench team for providing the original annotations and evaluation framework.

## License

This benchmark is released for **research purposes only**.