K-STR-Bench

K-STR-Bench is a unified, challenge-aware benchmark suite for Korean scene text recognition (STR).

It combines three complementary components: AIHub-Standardized, a reproducible large-scale STR resource with nearly 900K valid Korean text instances; KAIST-Corrected, a reconstructed and corrected legacy Korean scene-text dataset; and GIST-Signboard, a newly collected real-world signboard dataset covering challenging outdoor conditions.

The three components are organized under a common Korean-only cropped-STR protocol, supporting reproducible training, cross-dataset evaluation, and challenge-wise robustness analysis.

Code and instructions are available in the GitHub repository.

Myounghun Han, Jin-Hyuk Hong. ACCV 2026.

Subsets

Subset	Role	Scenes	Text	Korean-only	Train	Val	Test
AIHub_Standardized	Train, val, and in-domain test	533,469	1,937,042	897,055	717,669	89,716	89,670
KAIST_Corrected	Evaluation only	2,484	7,516	4,008	—	—	4,008
GIST_Signboard	Evaluation only	5,185	34,687	16,078	—	—	16,078

Korean-only counts are the samples used for K-STR-Bench. Train, val, and test for AIHub_Standardized are that Korean-only split. KAIST_Corrected and GIST_Signboard are test only.

GIST_Signboard and KAIST_Corrected also carry challenge tags on the Korean-only test samples. A sample may have more than one tag.

Subset	Artistic	Curve	Sailent	Incomplete	Multi-oriented	Low-visibility
KAIST_Corrected	563	128	446	56	411	282
GIST_Signboard	2,511	351	2,795	711	4,189	1,070

What this repository contains

GIST_Signboard and KAIST_Corrected images and label tables are included.

AIHub images and the original AIHub annotation files are not included. This repository has the K-STR-Bench metadata, splits, and reviewed transcriptions. Rebuild AIHub_Standardized locally from the AIHub dataset 야외 실제 촬영 한글 이미지.

Use these label files:

Subset	File
AIHub_Standardized	text.v2.0.csv
KAIST_Corrected	text.v2.0.csv
GIST_Signboard	text.v1.0.csv

Each text sample is keyed by text_id. A row points back to its scene with scene_id.

Evaluation

Score matched predictions with word accuracy and jaso-level 1-NED. On GIST_Signboard and KAIST_Corrected, report the same two scores for each challenge tag.

The scoring command and file format are documented in the GitHub repository.

License

This benchmark does not use one license for every file.

Resource	Terms
GIST_Signboard images and annotations	CC BY 4.0
K-STR-Bench metadata, ids, splits, and author-written labels	CC BY 4.0
KAIST_Corrected	CC BY-SA 3.0
Original AIHub images and annotations	Not redistributed. AIHub Terms of Use

CC BY 4.0 covers only material created and released by the K-STR-Bench authors. It does not cover the original AIHub dataset.

Citation

@inproceedings{han2026kstrbench,
    title     = {K-STR-Bench: A Challenge-Aware Benchmark Suite for Korean Scene Text Recognition},
    author    = {Han, Myounghun and Hong, Jin-Hyuk},
    booktitle = {Asian Conference on Computer Vision (ACCV)},
    year      = {2026}
}

Also cite the original AIHub and KAIST datasets when using those subsets.