SEA-DocScene: A Multi-Scene and Multi-Category Real-World OCR Benchmark

Dataset status: Preview page — data files have not been released. Coming soon. No dataset files, samples, or downloadable archives are distributed through this repository at this stage.

This dataset is developed and maintained by the [Yunnan Key Laboratory of Artificial Intelligence] at [Kunming University of Science and Technology].

This repository currently serves as the official project and documentation page for the dataset.

Overview

We present a document OCR and layout analysis benchmark covering five Southeast Asian languages: Khmer (km), Lao (lo), Burmese/Myanmar (my), Thai (th) and Vietnamese (vi). The benchmark is designed to evaluate text detection, text recognition and document layout analysis on real-world, in-the-wild documents from these low-resource script communities, spanning printed documents, scanned pages, presentation slides and photographic scene images.

Each page is annotated with page-level layout regions. Every region is described by a polygon, a layout category, a reading order and, for text-bearing regions, the transcript text plus line-level spans. Page metadata records image dimensions, the document source category, the language and the page layout style. The benchmark is intended to support both end-to-end OCR systems and document layout understanding models, and to expose the gap between high-resource and low-resource scripts.

Planned contents may include:

• Document page images (scanned pages, rendered documents and photographic scene images)
• Page-level layout annotations with polygons, categories and reading order
• Text transcripts at region level and line-span level
• A held-out benchmark split per language

Languages Covered

Code	Language	Script
km	Khmer	Khmer
lo	Lao	Lao
my	Burmese / Myanmar	Myanmar
th	Thai	Thai
vi	Vietnamese	Latin (with diacritics)

Document Categories

The benchmark covers 12 document source categories, each represented in every language:

Category	English
公文	Official / administrative documents
学术文献	Academic literature
研究报告	Research reports
书籍	Books
彩色教科书	Color textbooks
杂志	Magazines
报纸	Newspapers
试卷	Exam papers
PPT转PDF	Presentation slides (PPT exported to PDF)
笔记	Handwritten notes and note pages
证件照	ID / document photos
场景图像	Scene images (documents captured in the wild)

Annotation Format

Annotations are provided as JSON, one record per page. Each page record contains:

• layout_dets — list of layout regions, each with a polygon poly, a category_type, a reading order, an anno_id, optional text and line_with_spans (line-level text with polygons), and attribute (text language, background, rotation)
• page_info — page_no, page_code, document_code, image_path, width, height and page_attribute (data_source, language, layout, special_issue)
• extra — relation metadata between regions

Layout categories include: text_block, title, figure, table, header, footer, page_number, figure_caption, table_caption, equation_isolated, page_footnote, figure_footnote, table_footnote, equation_caption and reference.

Scale

The figures below reflect the dataset as currently prepared internally. Final released counts may change.

Language	Pages	Document categories
km	1,826	12
lo	2,016	12
my	1,844	12
th	1,755	12
vi	2,137	12
Total (benchmark split)	9,578	—

Availability

The dataset is currently being prepared for release.

Researchers interested in the dataset, potential collaboration, or release notifications may contact:

• Laboratory: [Yunnan Key Laboratory of Artificial Intelligence]
• Institution: [Kunming University of Science and Technology ]
• Email: [maocunli@163.com]
• Principal investigator: [Cunli Mao]

When contacting us, please include your name, affiliation, intended research use, and the part of the dataset you are interested in.

Access Policy

The final access mechanism and license are under review. Publication of this page does not grant access to the dataset or permission to use any underlying data.

The dataset may ultimately be released as:

• [Publicly downloadable / gated access / application-based access]
• [Research-only / non-commercial / another policy, if already approved]

Final terms will be published on this page before data files become available.

Responsible Use

The dataset contains real documents collected from public and institutional sources and may include personal identifiers, official seals, signatures, contact details or other sensitive content. Some documents may reflect the views or biases of their original authors, which are not endorsed by the maintainers. Users are expected to avoid attempting to identify individuals, and to respect privacy and applicable data protection regulations.

Users will be expected to comply with the final license, applicable laws, ethical requirements, and any institutional restrictions described at release time.

Citation

Citation information will be added when the accompanying paper or technical report becomes available.

Version History

• 2026-09-18 — Initial preview page created; no data files released.