union14m_l_cropped

A WebDataset-format mirror of the training portion of Union14M-L — 3.23 M already-cropped scene-text-recognition samples drawn from 14 public STR datasets.

This repo only contains the training splits (easy, normal, medium, hard, challenging); the bundled val split and the Union14M-Benchmarks/ evaluation suite are not included. The crops are the original axis-aligned-bbox crops produced by the Union14M authors — we just repackage them.

Stats

• Samples: 3,230,742
• Shards: 647 (shard-{00000..00646}.tar)
• Format: webdataset
• On-disk size: ~20 GB

Split breakdown (from upstream jsonls):

split	samples
easy	2,076,161
normal	218,154
medium	145,525
hard	308,025
challenging	482,877
total	3,230,742

Per-sample split information is not preserved in this repackaging; if you need it, refer to the upstream jsonl files (the original filename is encoded in the JPEG's content, not in our key).

Layout

data/shard-00000.tar
    union14m-00000-00000000.jpg
    union14m-00000-00000000.txt
    ...

Loading

import webdataset as wds

url = "https://huggingface.co/datasets/nagohachi/union14m_l_cropped/resolve/main/data/shard-{00000..00646}.tar"
ds = wds.WebDataset(url).decode("pil").to_tuple("jpg", "txt")
for image, text in ds:
    ...

Preprocessing

The shards were produced by streaming the official 12 GB Union14M-L.tar.gz once and joining each full_images/<source>/imgs/*.jpg member with the union of the five train_annos/mmocr0.x/train_*.jsonl files. Validation images and benchmark images are skipped.

Script: src/preproc/image/union14m_str.py in nagohachi/PetitMLLM (the calling project repo).

License

The repackaging code and the curation of Union14M itself are released under the MIT License (Copyright © 2023 Qing Jiang, per the Union14M repository LICENSE).

The underlying images are aggregated from 14 distinct source datasets each with their own license. The MIT license on this repo applies to the repackaging only — please consult the original dataset for the licensing terms of any image you intend to use, especially for commercial purposes.

Source datasets (per the Union14M source_dataset doc):

1. KAIST (2011)
2. NEOCR (2011)
3. Uber-Text (2017)
4. RCTW (2017)
5. IIIT-ILST (2017)
6. MTWI (2018)
7. COCOTextV2 (2018)
8. LSVT (2019)
9. MLT19 (2019)
10. ReCTS (2019)
11. ArT (2019)
12. IntelOCR / OpenVINO (2021)
13. TextOCR (2021)
14. HierText (2022)

Several of these datasets are research-only or carry attribution / non-commercial clauses — verify before commercial use.

Citation

If you use this data, cite the original Union14M paper:

@article{jiang2023revisiting,
  title={Revisiting Scene Text Recognition: A Data Perspective},
  author={Jiang, Qing and Wang, Jiapeng and Peng, Dezhi and Liu, Chongyu and Jin, Lianwen},
  journal={ICCV},
  year={2023}
}