STIPLAR: Scene Text Image Pairs of Low-resource lAnguage and Real-world data

STIPLAR is a real-world scene text image dataset containing Korean, Arabic, and Japanese text image pairs collected from MLT-2019 and web sources, designed for fine-tuning STELLAR on low-resource languages.

To do

Dataset download location will be updated.

Dataset

For Stage 2 training, we utilize STIPLAR, our newly proposed scene text image pairs of low-resource language and real-world data. The dataset for each language can be downloaded from Link.

Statistics

Version 1 (2025.08.06)

Lang.	Type	Train Open	Train Crawl	Train Total	Eval Open	Eval Crawl	Eval Total
Korean	Full image	269	317	586	68	80	148
Korean	Text image pair	1456	6229	7685	362	1717	2079
Arabic	Full image	252	56	308	64	15	79
Arabic	Text image pair	1879	3460	5339	450	531	981
Japanese	Full image	97	252	349	25	63	88
Japanese	Text image pair	356	1282	1638	97	288	385

Train Open, Eval Open: Images from Open-source dataset
Train Crawl, Eval Crawl: Images from web crawling.

Version 2 (2025.08.15)

In the Arabic Train Crawl, 1 Full image and 3 Text image pairs have been removed.

Lang.	Type	Train Open	Train Crawl	Train Total	Eval Open	Eval Crawl	Eval Total
Korean	Full image	269	317	586	68	80	148
Korean	Text image pair	1456	6229	7685	362	1717	2079
Arabic	Full image	252	55	307	64	15	79
Arabic	Text image pair	1879	3457	5336	450	531	981
Japanese	Full image	97	252	349	25	63	88
Japanese	Text image pair	356	1282	1638	97	288	385

Train Open, Eval Open: Images from Open-source dataset
Train Crawl, Eval Crawl: Images from web crawling.

Version 3 (2025.08.24)

In the Arabic Train Open, 1 Full image and 1 Text image pairs have been removed.

Lang.	Type	Train Open	Train Crawl	Train Total	Eval Open	Eval Crawl	Eval Total
Korean	Full image	269	317	586	68	80	148
Korean	Text image pair	1456	6229	7685	362	1717	2079
Arabic	Full image	251	55	306	64	15	79
Arabic	Text image pair	1878	3457	5335	450	531	981
Japanese	Full image	97	252	349	25	63	88
Japanese	Text image pair	356	1282	1638	97	288	385

Train Open, Eval Open: Images from Open-source dataset
Train Crawl, Eval Crawl: Images from web crawling.

Evaluation

Download the STIPLAR-eval dataset from Link and unzip the files. This is a newly re-numbered version that combines the eval-crawling folder and the eval-mlt2019 folder from the STIPILAR dataset.

├── anonymous-stellar/
│   ├── STIPLAR-ko-eval/
│   ├── STIPLAR-ar-eval/
│   └── STIPLAR-jp-eval/
│       ├── i_s/
│       ├── t_f/
│       └── i_full/

License

Dataset

The dataset contained in this repository consists of web-crawled images originally distributed under the following open licenses: CC0, CC BY, CC BY-NC, and PDM.

Each image retains its original license, which is specified in the metadata.
For CC BY images, proper attribution to the original creator is required.
For CC BY-NC images, commercial use is not permitted.
For CC0 and PDM images, the content is in the public domain and may be freely used.

The dataset curators do not claim ownership of the original images, only of the metadata and dataset organization, which are released under CC BY 4.0.