Soxavin/ardb-layout-coco-v5

ARDB Daily Bulletin Layout Detection (COCO)

A document-layout object-detection dataset built from ARDB (Agricultural and Rural Development Bank of Cambodia) daily market-price bulletins — born-digital Khmer-language PDFs, one price table per page. Every page is annotated with bounding boxes for five layout regions (table, text, section header, page furniture, picture).

Each row is a full page image with its box annotations embedded, so the Dataset Viewer shows the image beside its labels.

The source corpus spans a multi-year range (2022–2026), covering both structural bulletin templates in circulation over that period. Supersedes ardb-layout-coco-v4 (same annotations; that card documents the split-assignment issue this version fixes).

Dataset structure

Split	Pages	Boxes	Source documents
train	136	636	45 bulletins (14 retail_only + 31 wholesale_retail)
validation	9	42	3 bulletins (1 retail_only + 2 wholesale_retail)
test	21	98	7 bulletins (1 retail_only + 6 wholesale_retail)
total	166	776	55 bulletins

Splits are assigned by document, clustered by issue date (documents dated within 1 day of each other are kept in the same split) and stratified by template era (each era's clusters are split independently, then merged), so no bulletin's pages appear in more than one split, no near- duplicate adjacent-day bulletin straddles a split boundary, and both structural templates are represented in every split.

Fields

Column	Type	Description
image	image	Embedded page image (JPG).
image_id	int64	Row index within the split.
file_name	string	Source image file name.
doc_id	string	Identifier of the originating bulletin.
source	string	Originating PDF (provenance).
width	int64	Image width in pixels.
height	int64	Image height in pixels.
objects	struct of lists	Per-box annotations (see below).

objects is a struct of parallel lists carrying the full COCO annotation fields:

• id — annotation id (unique within its split; renumbered from v4 since pages moved between splits — not semantically meaningful, no consumer reads it)
• bbox — [x, y, width, height] in pixels (COCO convention)
• category_id — class index (0–4, see below)
• category — human-readable class name
• area — box area in px²
• iscrowd — always 0
• score — annotation confidence (1.0, human-verified)

Classes

id	name	Notes
0	Table	The price table — exactly one per page, covering the full table including header row and label columns.
1	Text	Body / paragraph text, including footer notes.
2	Section-Header	Headings and titles.
3	Page-Furniture	Page headers and footers (2 per page).
4	Picture	Logos and stamps (1 per page).

Usage

from datasets import load_dataset

ds = load_dataset("Soxavin/ardb-layout-coco-v5")   # splits: train / validation / test
row = ds["train"][0]
row["image"]                  # PIL image
row["objects"]["bbox"]        # list of [x, y, w, h]
row["objects"]["category"]    # list of class names

Data collection & annotation

1. Source. Publicly published ARDB daily market-price bulletins spanning 2022–2026, rendered page-by-page at 200 DPI.
2. Pre-annotation. Candidate boxes were generated automatically with a document-layout detector (Surya) and confidence-filtered, then mapped to the five-class set above.
3. Human correction. Every page was reviewed and corrected in Roboflow: fragmented table regions merged to one box per page, footer text unified under a single Text label, box edges tightened, mislabels fixed, and spurious boxes removed. All final annotations carry score = 1.0.
4. Splitting. Splits are assigned by document, clustered by issue date to avoid adjacent-day near-duplicate leakage, and stratified by template era so both structural templates are represented in every split. Assignment is deterministic.

Limitations

• Two structural templates. The corpus spans two distinct bulletin layouts (a retail-only 6-column table pre-May 2024, and a combined wholesale/retail 9-column table from May 2024 onward). Both are represented in every split, though not in exactly proportional counts — cluster integrity (no adjacent-day leakage) takes priority over exact proportionality.
• Small scale. 166 pages / 55 documents; suitable for fine-tuning and evaluation on this document family, not as a general-purpose layout corpus.
• Minority classes. Section-Header and Text have substantially fewer boxes than the other three classes, reflecting their genuine lower frequency per page (not every page has a section header or footer text), rather than an annotation gap.

License

The annotations in this dataset (bounding boxes, class labels, and metadata) are released under CC BY-NC 4.0 — free to use for non-commercial research with attribution. The page images are reproductions of bulletins published by ARDB; they are included for research use only, and users are responsible for complying with the source documents' terms. This dataset is not affiliated with or endorsed by ARDB.