Hugging Face's logo

Datasets:
PaddlePaddle
/
Real5-OmniDocBench

Modalities:
Image
Dataset card Data Studio Files
xet
Community
Real5-OmniDocBench / README.md
zhouchangda
update mertic
cc119ae
preview code
|
raw
history blame contribute delete
23 kB

Real5-OmniDocBench

Real5-OmniDocBench is a brand-new benchmark oriented toward real-world scenarios, which we constructed based on the OmniDocBench v1.5 dataset. The dataset comprises five distinct scenarios: Scanning, Warping, Screen-Photography, Illumination, and Skew. Apart from the Scanning category, all images were manually acquired via handheld mobile devices to closely simulate real-world conditions. Each subset maintains a one-to-one correspondence with the original OmniDocBench, strictly adhering to its ground-truth annotations and evaluation protocols. Given its empirical and realistic nature, this dataset serves as a rigorous benchmark for assessing the robustness of document parsing models in practical applications.

Key Features

1. Real-world Scenarios

Real5-OmniDocBench contains five challenging and representative photographic scenarios:

  • Scanning: Images captured by scanning devices, simulating flat and clean document scans.
  • Warping: Photos of documents with visible page curvatures, mimicking distortions caused by book spines or folding.
  • Skew: Documents photographed at an angle, resulting in perspective distortion.
  • Screen-Photography: Photographs of screens displaying documents, introducing moiré patterns and reflections.
  • Illumination: Images taken under varied lighting conditions, including shadows and glare.

2. Comprehensive Coverage

For each scenario, the dataset contains 1,355 images, corresponding exactly to the original pages in the OmniDocBench evaluation set. This enables direct, controlled comparison of model performance across different real-world conditions.

3. High-Quality Realism

All images are captured manually under real-world conditions. Care was taken to cover typical distortions and artifacts encountered in mobile or camera-based document digitization.

4. Evaluation Protocol Compatibility

The evaluation scripts and metrics from OmniDocBench (e.g., Normalized Edit Distance, BLEU, METEOR, TEDS, COCODet) are fully compatible and directly applicable to Real5-OmniDocBench.

Dataset Structure

The dataset follows the directory structure shown below:

Real5-OmniDocBench/
├── Real5-OmniDocBench-Warping/
│   ├── book_en_[搬书匠#20][HTML5 Canvas].2011.英文版_page_208.png
│   └── ...
├── Real5-OmniDocBench-Illumination/
│   └── ...
├── Real5-OmniDocBench-Scanning/
│   └── ...
├── Real5-OmniDocBench-Screen-Photography/
│   └── ...
└── Real5-OmniDocBench-Skew/
    └── ...

Usage & Evaluation

  • Input: For each scenario, models should take the corresponding set of 1,355 images as input.
  • Output: Model predictions should follow the same format as OmniDocBench, e.g., Markdown for end-to-end parsing.
  • Evaluation: Use the official OmniDocBench evaluation scripts and metrics for assessment. Direct comparison across scenarios is encouraged to measure model robustness.

Benchmark Results

To illustrate the impact of real-world visual distortions on document parsing, we evaluated several mainstream models across all five scenarios in Real5-OmniDocBench. The same metrics as OmniDocBench are used: Overall↑, TextEdit↓, FormulaCDM↑, TableTEDS↑, and Reading OrderEdit↓.

1. Scanning

Model Type Methods Parameters Overall↑ TextEdit↓ FormulaCDM↑ TableTEDS↑ Reading OrderEdit↓
Pipeline Tools Maker-1.8.2 - 70.27 0.223 77.03 56.05 0.238
PP-StructureV3 - 84.68 0.094 84.34 79.06 0.092
General VLMs GPT-5.2 - 84.43 0.142 85.68 81.78 0.109
Qwen2.5-VL-72B 72B 86.19 0.110 86.14 83.41 0.114
Gemini-2.5 Pro - 89.25 0.073 87.44 87.62 0.098
Qwen3-VL-235B-A22B-Instruct 235B 89.43 0.059 89.01 85.19 0.066
Gemini-3 Pro - 89.47 0.071 88.16 87.37 0.078
Specialized VLMs Dolphin 322M 72.16 0.154 64.58 67.27 0.130
Dolphin-1.5 0.3B 83.39 0.097 76.25 83.65 0.090
MinerU2-VLM 0.9B 83.60 0.094 79.76 80.44 0.091
MonkeyOCR-pro-1.2B 1.9B 84.64 0.123 84.17 82.13 0.145
MonkeyOCR-3B 3.7B 84.65 0.100 84.16 79.81 0.143
Nanonets-OCR-s 3B 85.52 0.106 88.09 79.11 0.106
Deepseek-OCR 3B 86.17 0.078 83.59 82.69 0.085
dots.ocr 3B 86.87 0.083 83.27 85.68 0.081
MonkeyOCR-pro-3B 3.7B 86.94 0.103 86.29 84.86 0.141
MinerU2.5 1.2B 90.06 0.052 88.22 87.16 0.050
PaddleOCR-VL 0.9B 92.11 0.039 90.35 89.90 0.048
PaddleOCR-VL-1.5 0.9B 93.43 0.037 93.04 90.97 0.045

In the scanning scenario, the models’ overall performance remains high, with PaddleOCR-VL-1.5 typically achieving the best results.

2. Warping

Model Type Methods Parameters Overall↑ TextEdit↓ FormulaCDM↑ TableTEDS↑ Reading OrderEdit↓
Pipeline Tools Maker-1.8.2 - 58.98 0.349 72.71 39.08 0.390
PP-StructureV3 - 59.34 0.376 68.22 47.40 0.261
General VLMs GPT-5.2 - 76.26 0.239 80.90 71.80 0.165
Gemini-2.5 Pro - 87.63 0.092 86.50 85.59 0.109
Qwen2.5-VL-72B 72B 87.77 0.086 88.85 83.06 0.102
Gemini-3 Pro - 88.90 0.086 88.10 87.20 0.087
Qwen3-VL-235B-A22B-Instruct 235B 89.99 0.051 89.06 85.95 0.064
Specialized VLMs Dolphin-1.5 0.3B 50.50 0.383 47.24 42.52 0.309
Dolphin 322M 60.35 0.316 61.06 51.58 0.247
Deepseek-OCR 3B 67.20 0.328 73.59 60.80 0.226
MinerU2-VLM 0.9B 73.73 0.202 77.72 63.65 0.173
MonkeyOCR-pro-1.2B 1.9B 76.59 0.196 78.85 70.52 0.221
MonkeyOCR-3B 3.7B 77.27 0.164 79.08 69.18 0.211
MonkeyOCR-pro-3B 3.7B 78.90 0.168 79.55 73.94 0.212
Nanonets-OCR-s 3B 83.56 0.121 86.24 76.57 0.124
MinerU2.5 1.2B 83.76 0.154 85.92 80.71 0.104
PaddleOCR-VL 0.9B 85.97 0.093 85.45 81.77 0.092
dots.ocr 3B 86.01 0.087 85.03 81.74 0.093
PaddleOCR-VL-1.5 0.9B 91.25 0.053 90.94 88.10 0.063

Warping documents introduce moderate challenges, causing a slight performance drop, but specialized VLMs maintain competitive accuracy.

3. Screen-Photography

Model Type Methods Parameters Overall↑ TextEdit↓ FormulaCDM↑ TableTEDS↑ Reading OrderEdit↓
Pipeline Tools Maker-1.8.2 - 63.65 0.290 72.73 47.21 0.325
PP-StructureV3 - 66.89 0.204 73.26 47.82 0.165
General VLMs GPT-5.2 - 76.75 0.208 79.27 71.73 0.148
Qwen2.5-VL-72B 72B 86.48 0.100 87.46 82.00 0.102
Gemini-2.5 Pro - 87.11 0.103 85.30 86.31 0.117
Gemini-3 Pro - 88.86 0.084 87.33 87.65 0.087
Qwen3-VL-235B-A22B-Instruct 235B 89.27 0.068 88.72 85.85 0.071
Specialized VLMs Dolphin 322M 64.29 0.232 58.66 57.38 0.195
Dolphin-1.5 0.3B 69.76 0.205 61.80 68.00 0.177
Deepseek-OCR 3B 75.31 0.220 77.68 70.26 0.169
MinerU2-VLM 0.9B 78.77 0.139 79.02 71.17 0.123
MonkeyOCR-pro-1.2B 1.9B 80.24 0.148 80.78 74.74 0.179
MonkeyOCR-3B 3.7B 80.71 0.122 81.33 73.04 0.177
MonkeyOCR-pro-3B 3.7B 82.44 0.124 81.55 78.13 0.177
PaddleOCR-VL 0.9B 82.54 0.103 83.58 74.36 0.107
Nanonets-OCR-s 3B 84.86 0.112 86.65 79.09 0.117
dots.ocr 3B 87.18 0.081 85.34 84.26 0.079
MinerU2.5 1.2B 89.41 0.062 87.55 86.83 0.053
PaddleOCR-VL-1.5 0.9B 91.76 0.050 90.88 89.38 0.059

4. Illumination

Model Type Methods Parameters Overall↑ TextEdit↓ FormulaCDM↑ TableTEDS↑ Reading OrderEdit↓
Pipeline Tools Maker-1.8.2 - 66.31 0.259 74.80 50.03 0.337
PP-StructureV3 - 73.38 0.158 77.75 58.19 0.126
General VLMs GPT-5.2 - 80.88 0.191 84.41 77.37 0.134
Qwen2.5-VL-72B 72B 87.25 0.087 86.44 84.03 0.097
Gemini-2.5 Pro - 87.97 0.083 86.13 86.11 0.103
Qwen3-VL-235B-A22B-Instruct 235B 89.27 0.060 87.81 86.05 0.070
Gemini-3 Pro - 89.53 0.073 87.78 88.14 0.080
Specialized VLMs Dolphin 322M 67.29 0.197 61.42 60.10 0.173
Dolphin-1.5 0.3B 75.61 0.159 70.04 72.69 0.133
Deepseek-OCR 3B 78.10 0.192 81.71 71.81 0.156
MinerU2-VLM 0.9B 80.51 0.135 80.72 74.29 0.123
MonkeyOCR-pro-1.2B 1.9B 82.11 0.144 82.07 78.67 0.172
MonkeyOCR-3B 3.7B 83.16 0.118 83.63 77.62 0.168
MonkeyOCR-pro-3B 3.7B 84.71 0.120 84.13 82.02 0.171
Nanonets-OCR-s 3B 85.01 0.099 87.94 76.96 0.112
dots.ocr 3B 87.57 0.068 85.07 84.44 0.076
MinerU2.5 1.2B 89.57 0.065 88.36 86.87 0.062
PaddleOCR-VL 0.9B 89.61 0.049 86.66 87.02 0.055
PaddleOCR-VL-1.5 0.9B 92.16 0.046 91.80 89.33 0.051

Severe illumination variations present significant challenges, resulting in a noticeable performance drop for most models.

5. Skew

Model Type Methods Parameters Overall↑ TextEdit↓ FormulaCDM↑ TableTEDS↑ Reading OrderEdit↓
Pipeline Tools PP-StructureV3 - 37.98 0.557 44.37 25.27 0.417
Maker-1.8.2 - 41.27 0.536 60.16 17.23 0.543
General VLMs GPT-5.2 - 75.00 0.257 80.27 70.47 0.167
Qwen3-VL-235B-A22B-Instruct 235B 86.56 0.077 83.96 83.41 0.091
Qwen2.5-VL-72B 72B 86.90 0.077 87.26 81.14 0.091
Gemini-2.5 Pro - 89.07 0.077 87.89 86.99 0.104
Gemini-3 Pro - 89.45 0.080 88.33 88.06 0.092
Specialized VLMs Dolphin-1.5 0.3B 28.16 0.553 25.60 14.18 0.419
Dolphin 322M 44.83 0.500 51.34 33.22 0.321
MonkeyOCR-pro-1.2B 1.9B 62.18 0.292 66.25 49.46 0.317
Deepseek-OCR 3B 63.01 0.327 73.27 48.48 0.231
MonkeyOCR-pro-3B 3.7B 64.47 0.251 69.06 49.42 0.301
MonkeyOCR-3B 3.7B 65.67 0.248 69.23 52.59 0.300
MinerU2-VLM 0.9B 68.16 0.230 74.45 53.07 0.191
MinerU2.5 1.2B 75.24 0.305 81.78 74.39 0.151
PaddleOCR-VL 0.9B 77.47 0.192 78.81 72.83 0.193
Nanonets-OCR-s 3B 81.98 0.121 85.78 72.22 0.133
dots.ocr 3B 84.27 0.087 85.73 75.74 0.094
PaddleOCR-VL-1.5 0.9B 91.66 0.047 91.00 88.69 0.061

Skewed documents also impact accuracy, but top-performing models still provide reasonable results.

Acknowledgements

Real5-OmniDocBench is based on the original OmniDocBench dataset and adopts its evaluation protocols. We thank the authors of OmniDocBench for their foundational work.

Citation

If you use Real5-OmniDocBench in your research, please cite our dataset and refer also to the original OmniDocBench paper.