Yesianrohn/STR-Synth

---
language:
- en
license: apache-2.0
size_categories:
- 10M<n<100M
task_categories:
- image-to-text
tags:
- ocr
- Scene Text Recognition
- synthetic data
- lmdb
- computer vision
---

# STR-Synth

[**Paper**](https://huggingface.co/papers/2602.06450) | [**GitHub**](https://github.com/YesianRohn/UnionST)

This repository serves as the supplementary dataset resource for the paper *What’s Wrong with Synthetic Data for Scene Text Recognition? A Strong Synthetic Engine with Diverse Simulations and Self-Evolution*, dedicated to summarizing and providing the representative synthetic datasets for Scene Text Recognition (STR) used in the paper's comparative experiments.

The dataset collection in this repository aggregates **6 mainstream STR synthetic datasets**: MJ, ST, SynthAdd, CurvedST, SynthTIGER, and UnrealText, with a total of **46 million samples** in total, covering a diverse range of scene text simulation scenarios and characteristics to support comprehensive comparative research on STR synthetic data performance.

In line with industry standards, all datasets in this repository are provided in the **lmdb file format**—the de facto standard adopted by the mainstream STR research protocol, ensuring seamless compatibility with most existing STR training and evaluation frameworks for straightforward integration and usage.

For the self-developed high-performance synthetic STR dataset proposed in the paper, **UnionST** (featuring superior diversity, label accuracy and cost-efficiency), please refer to its dedicated repository at: https://huggingface.co/datasets/Yesianrohn/UnionST.

## Citation

```bibtex
@inproceedings{ye2026wrong,
title={What's Wrong with Synthetic Data for Scene Text Recognition? A Strong Synthetic Engine with Diverse Simulations and Self-Evolution},
author={Ye, Xingsong and Du, Yongkun and Zhang, JiaXin and Li, Chen and LYU, Jing and Chen, Zhineng},
booktitle={CVPR},
year={2026}
}
```