Reproduce “SyntheticDoc: A Large Synthetic Dataset for Document Unwarping and Illumination Correction” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Daniel Woortmann, Tanguy Magne, Olga Sorkine-Hornung

SyntheticDoc generates synthetic training data for document unwarping and illumination correction by simulating bent paper and rendering it with varied documents, materials, cameras, and lighting, while retaining annotations. A lightweight baseline trained on this dataset leads methods on all five unwarping metrics on DocUNet, including aligned distortion of 0.252 versus 0.277 for AADD; the comparison does not isolate dataset effects.

Key insights

  • Training-data scale matters: increasing the dataset from 100,000 to 500,000 and then 1,000,000 samples steadily improves unwarping quality across visual distortion and OCR metrics.
  • High-resolution supervision improves visual reconstruction, but better image similarity does not automatically yield better OCR, underscoring the difference between geometric/photometric fidelity and text preservation.
  • A carefully simulated dataset can make a relatively simple baseline competitive with or better than more specialized methods, while synthetic-to-real gaps, storage demands, and limited camera conditions remain constraints.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.