Dataset · PaddlePaddle
Distorted document parsing benchmark
PaddlePaddle/Real5-OmniDocBench
Document image set with physical distortions to test OCR and parsing system robustness.
What's inside
Contains 6,775 document images in English and Chinese. These are 1,355 pages from OmniDocBench v1.5 recreated under five real capture conditions: scanning, warping, screen photography, varied lighting, and tilt.
What you can do with it
- Evaluating OCR and text extraction model robustness against real capture defects
- Testing document parsing VLM systems in challenging conditions
- Comparing alignment and preprocessing algorithms for scans and photos
- Tasks
- Image → text
- Languages
- Chinese, English
- Size
- 1K–10K rows
- Downloads this month
- 21.5K
- Likes
- 43
How to load it
from datasets import load_dataset
ds = load_dataset("PaddlePaddle/Real5-OmniDocBench")