SyntheticDoc: A Large Synthetic Dataset for Document Unwarping and Illumination Correction

Daniel Woortmann*, Tanguy Magne*, Olga Sorkine-Hornung
* Equal contribution

This repository contains the code and data for the ECCV paper "SyntheticDoc: A Large Synthetic Dataset for Document Unwarping and Illumination Correction".

📁 Dataset

Our dataset is available here. Note that, for now, only the rendered image, albedo, shading, UV, backward mapping, and metadata are available for each sample. The remaining annotations (normal maps and 3D coordinates) will be released soon.

Documentation about the currently released dataset can be found here.

The assets it was generated from (meshes, document pages and background materials) are available here.

💻 Code

The code to generate the dataset is available in the generation folder: generation/simulation simulates the deformed paper meshes, generation/rendering renders the final samples (images and annotations), and generation/backward_mapping inverts the rendered UV maps into the backward maps used to supervise unwarping.

The code to train the model and to run it on your own images is available in the training folder.

🪪 Citation

@inproceedings{SyntheticDoc:2026,
    author = {Woortmann, Daniel and Magne, Tanguy and Sorkine-Hornung, Olga},
    title = {{SyntheticDoc}: A Large Synthetic Dataset for Document Unwarping and Illumination Correction},
    booktitle={Computer Vision -- ECCV 2026},
    year = {2026},
}