Spatial1ntelligence/synthtext-recovered-mirror

SynthText — Recovered Mirror

This repository re-hosts the SynthText dataset introduced in the paper below, using the original generation code released by the authors and the original background/ground-truth data from its Academic Torrents distribution.

Ankush Gupta, Andrea Vedaldi, Andrew Zisserman. "Synthetic Data for Text Localisation in Natural Images." CVPR 2016. Original official page: https://www.robots.ox.ac.uk/~vgg/data/scenetext/ (download links are no longer available there)

Why this repository exists

The Oxford VGG lab's official page currently states only that "the download links for the SynthText dataset are no longer available from this website," without giving a reason. However, the authors' generation code (GitHub ankush-me/SynthText, Apache-2.0) and the original background images, depth maps, segmentation masks, and a pre-generated copy of the dataset remain available via Academic Torrents. This repository is that content, re-hosted as-is.

⚠️ License — please read

• The generation code is Apache License 2.0 and may be freely used.
• The background images, depth/segmentation data, and the generated SynthText dataset itself are not owned by the SynthText authors (the original repository's README states "We do not own the copyright to these images") and are governed by the terms stated on the Academic Torrents distribution page for this dataset, quoted here verbatim:
  • "Researcher shall use the Database only for non-commercial research and educational purposes."
  • "Researcher's employer shall also be bound by these terms."
  • Commercial use requires separate permission from Oxford University Innovation (the university's technology transfer office).
  • The University of Oxford may terminate the Researcher's access to the Database at any time.
• In short: use this dataset for non-commercial research and educational purposes only. Confirm the original license terms yourself before any commercial use.
• Copyright in the individual background images is distributed across many original sources that even the SynthText authors do not fully track. This repository simply re-hosts the content of the original Academic Torrents distribution as-is; it does not resolve the copyright status of the individual background images.

Contents

• SynthText.zip — approximately 800,000 synthetic scene-text images generated with the paper's code (identical to the original Academic Torrents distribution).
• The generation code itself is not included in this repository. See https://github.com/ankush-me/SynthText if you need it.

Citation

@InProceedings{Gupta16,
  author       = "Gupta, A. and Vedaldi, A. and Zisserman, A.",
  title        = "Synthetic Data for Text Localisation in Natural Images",
  booktitle    = "IEEE Conference on Computer Vision and Pattern Recognition",
  year         = "2016",
}

Sources

• Original paper / official page (downloads currently unavailable): https://www.robots.ox.ac.uk/~vgg/data/scenetext/
• Generation code: https://github.com/ankush-me/SynthText
• Original data (Academic Torrents): https://academictorrents.com/details/2dba9518166cbd141534cbf381aa3e99a087e83c