README.md

metadata

license: other
task_categories:
  - image-to-text
language:
  - ko
  - en
tags:
  - scene-text-recognition
  - ocr
pretty_name: >-
  An Atomic Multi-Tag Taxonomy for Fine-Grained Diagnosis of Scene Text
  Recognition Models
size_categories:
  - 10K<n<100K
configs:
  - config_name: K_STR_Bench
    data_files:
      - split: test
        path: K_STR_Bench/text.v1.0.csv
    default: true
  - config_name: Union14M_Benchmark
    data_files:
      - split: test
        path: Union14M_Benchmark/text.v1.0.csv
  - config_name: Union14M_General
    data_files:
      - split: test
        path: Union14M_General/text_v.1.0.csv

An Atomic Multi-Tag Taxonomy for Fine-Grained Diagnosis of Scene Text Recognition Models

Atomic multi-tag annotations for K-STR-Bench and Union14M. Each text crop may contain multiple visual challenge tags, enabling fine-grained analysis of recognition performance under individual and co-occurring challenges. Images are stored once. Tags are columns in the CSV. image_path is relative to the folder that contains the CSV.

Myounghun Han, Jin-Hyuk Hong · WACV 2027

Accepted to WACV 2027. Code and documentation are on GitHub.

Data Sources

K-STR-Bench. A challenge-aware benchmark for Korean scene text recognition. GitHub | Hugging Face. The split here is K_STR_Bench, from GIST_Signboard and KAIST_Corrected. AIHub_Standardized is part of K-STR-Bench and is not included. The K-STR-Bench release includes every crop and the columns is_only_korean and is_illegible for filtering. This release is already filtered: text.v1.0.csv and the image folders contain only crops where is_only_korean is true and is_illegible is false.

Union14M. A scene text recognition dataset collected from 14 public datasets (Jiang et al., ICCV 2023). GitHub. Union14M-L and Union14M-Benchmark are released from that repository. The splits here are Union14M_Benchmark and Union14M_General. We annotate 7,888 samples from the six challenge-specific subsets of Union14M-Benchmark other than Incomplete, along with 9,994 randomly sampled images from its General subset. Each CSV and its image folder already contain only the crops in this release. Tag columns in these tables are separate from the original Union14M annotations.

Image folders are stored as text_images.tar. The CSV files stay beside those archives. Extracting an archive recreates a text_images folder next to it. This release does not include the inclusive and exclusive label files.

K_STR_Bench/
├── text.v1.0.csv
├── GIST_Signboard/text_images.tar
└── KAIST_Corrected/text_images.tar
Union14M_Benchmark/
├── text.v1.0.csv
└── text_images.tar
Union14M_General/
├── text_v.1.0.csv
└── text_images.tar

How to use

Download this dataset, then extract each text_images.tar. The CSV files are already unpacked. image_path is relative to the folder that contains the CSV. Create the inclusive and exclusive label files from the CSV with the code on GitHub.

git clone https://github.com/hoonisone/AtomicTags.git
cd AtomicTags
pip install -U huggingface_hub
hf download hoonisone/AtomicTags --repo-type dataset --local-dir data
python prepare.py --dataset-dir data --pipeline extract_images
python prepare.py --dataset-dir data --pipeline make_tag_labels

./K_STR_Bench/labels/inclusive/stylized.txt

0/gist_7_4.jpg	이글
0/gist_7_5.jpg	피플

./K_STR_Bench/labels/inclusive/stylized.jsonl

{"filename":"0/gist_7_4.jpg","text":"이글","width":87,"height":74}
{"filename":"0/gist_7_5.jpg","text":"피플","width":104,"height":202}

License

Source code is under the MIT License. That license does not cover the images or the annotations. The Atomic Tag annotations are under CC BY 4.0.

These files do not share one license. Union14M-Benchmark and Union14M-General are not separate licenses. An image keeps the license of the Union14M dataset it came from. Names and terms below follow the Union14M source table. Unknown means that table does not state a license.

Source	Subset	License / terms
K-STR-Bench	GIST_Signboard	CC BY 4.0
K-STR-Bench	KAIST_Corrected	CC BY-SA 3.0
K-STR-Bench	AIHub_Standardized	Not included. Not redistributed; AIHub Terms of Use
Union14M	KAIST	CC BY-SA 3.0
Union14M	NEOCR	CC BY-NC-SA 4.0
Union14M	Uber-Text	Unknown
Union14M	RCTW	Unknown
Union14M	IIIT-ILST	CC BY 4.0
Union14M	MTWI	CC BY-NC 4.0
Union14M	COCOTextV2	CC BY 4.0
Union14M	LSVT	Unknown
Union14M	MLT19	CC BY-NC 4.0
Union14M	ReCTS	Unknown
Union14M	ArT	Unknown
Union14M	IntelOCR	Apache License 2.0
Union14M	TextOCR	CC BY 4.0
Union14M	HierText	CC BY-SA 4.0

Statistics

Inclusive counts crops where that tag is true, including crops with other true tags. Exclusive counts crops whose only true tag is that tag.

Tag	K-STR-Bench		Union14M-General		Union14M-Benchmark
	Inclusive	Exclusive	Inclusive	Exclusive	Inclusive	Exclusive
Crops	20,086		9,994		7,888
Stylized	2,273	821	791	171	1,727	311
Inconsistent size	602	88	253	25	599	19
Incomplete	767	193	844	133	759	82
Low visibility	1,352	550	6,984	4,743	307	25
Noisy	2,844	1,466	696	109	1,540	143
Distracted	3,019	457	880	31	3,733	127
Perspective distortion	3,268	181	1,231	33	676	4
Non-single-line	118	34	0	0	0	0
Non-straight	479	60	155	6	2,731	203
Rotated	4,554	1,052	2,127	214	4,928	536

True tags on a crop	K-STR-Bench	Union14M-General	Union14M-Benchmark
0	9,889	1,389	745
1	4,902	5,465	1,450
2	2,560	1,685	2,671
3	1,892	852	2,100
4	676	467	738
5	135	115	156
6	27	20	21
7	3	1	6
8	2	0	1

Citation

If you use these annotations, please cite:

@inproceedings{atomicmultitag2027,
    title     = {An Atomic Multi-Tag Taxonomy for Fine-Grained Diagnosis of Scene Text Recognition Models},
    author    = {Han, MyoungHun and Hong, Jin-Hyuk},
    booktitle = {Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)},
    year      = {2027}
}