nakamura196/ndl-layout-dataset

NDL-DocL Kotenseki Layout Dataset (YOLO format)

A YOLO-formatted conversion of the kotenseki (pre-modern Japanese materials, 古典籍資料) subset of the NDL-DocL dataset published by the National Diet Library of Japan (NDL).

• Source dataset: https://github.com/ndl-lab/layout-dataset
• Source images: NDL Digital Collections https://dl.ndl.go.jp/

国立国会図書館が公開する NDL-DocL データセットのうち、古典籍資料を YOLO 形式（Ultralytics 互換）に変換したものです。

This is a modified/derived version. The bounding boxes were converted from the original Pascal VOC XML to normalized YOLO class cx cy w h text files, the images were downscaled, and a train/validation/test split was added. Neither the National Diet Library nor the original authors produced this derived form.

本データセットは編集・加工を行った派生物です。 原データ（Pascal VOC 形式の XML） から YOLO 形式への変換、画像の縮小、train/val/test 分割の付与を行っています。 これらの加工は国立国会図書館が行ったものではありません。

Dataset Details

• Original data curated by: National Diet Library, Japan (annotations created by NDL staff)
• Converted / curated by: Satoru Nakamura (中村覚), Historiographical Institute, The University of Tokyo
• Language: Japanese (classical / pre-modern, incl. kuzushiji cursive script)
• Source dataset version: NDL-DocL ver. 1.0 (released 2019-12-09; the only released version besides the 2019-11-08 beta). The 1,219 images here match the published image count of the kotenseki subset exactly.
• Derivative created: May 2024
• Task: document layout analysis as object detection
• Size: 1,219 images / 25,820 bounding boxes

Labels

Five classes, matching the label scheme of the NDL-DocL kotenseki subset. Note that the kindai (post-Meiji, 明治期以降刊行資料) subset of NDL-DocL uses a different label scheme (6_headline, 7_caption, 8_textline, 9_table) and is not included here.

YOLO id	Name	説明	Description
0	1_overall	資料範囲全体	The whole page / material region
1	2_handwritten	くずし字の文字ライン	Text line in cursive (kuzushiji) script
2	3_typography	楷書体・行書体の文字ライン	Text line in regular / semi-cursive script
3	4_illustration	イラスト	Illustration
4	5_stamp	印影（蔵書印等）	Seal impression (e.g. ownership seals)

Text lines are allowed to appear inside illustration regions (common in ukiyo-e), so 4_illustration boxes routinely overlap text-line boxes.

Class distribution

id	Name	train	val	test	total boxes	images containing it
0	1_overall	853	244	122	1,219	1,219
1	2_handwritten	9,811	2,656	1,384	13,851	721
2	3_typography	6,533	1,812	917	9,262	626
3	4_illustration	813	200	106	1,119	399
4	5_stamp	243	73	53	369	194
	total	18,253	4,985	2,582	25,820	1,219

The dataset is heavily imbalanced: 5_stamp accounts for 1.4% of all boxes and appears in only 194 of 1,219 images. Every image has exactly one 1_overall box, and no image has an empty label file.

Dataset Structure

data/
  data.yaml                 # Ultralytics config for the original page-level split
  data_pid.yaml             # Ultralytics config for the PID-level split (recommended)
  pid_{train,val,test}.txt  # image lists for the PID-level split
  pid_assignment.json       # PID -> split, for the PID-level split
  images/{train,val,test}/  # 853 / 244 / 122 JPEG images
    metadata.jsonl          # boxes, PID, page, size, pid_split (for 🤗 datasets)
  labels/{train,val,test}/  # one .txt per image, YOLO format
scripts/
  make_pid_split.py         # regenerates data_pid.yaml / pid_*.txt / pid_assignment.json
  make_metadata.py          # regenerates the metadata.jsonl files

The image and label files never move: both splits index the same 1,219 files. data.yaml is kept as published in May 2024 so that results reported against it remain reproducible.

Label files contain one box per line:

<class_id> <x_center> <y_center> <width> <height>

with all coordinates normalized to [0, 1] relative to the image size.

File names follow the NDL convention (PID)_(コマ番号), e.g. 2534020_0003 is frame 3 of https://dl.ndl.go.jp/pid/2534020. This makes it possible to look up the bibliographic record and the full-resolution image for every item.

Usage (Ultralytics)

from ultralytics import YOLO

model = YOLO("yolov8n.pt")
model.train(data="data/data.yaml", epochs=100, imgsz=640)

data/data.yaml uses paths relative to the data/ directory, so run training from there (or rewrite the train / val / test entries to absolute paths).

PID-level split (recommended for evaluation)

The original split is random per page, so pages of the same item land in different splits and test scores come out too high (see Limitations). data/data_pid.yaml provides a second split in which no PID crosses a split boundary:

	train	val	test	total
PIDs	67	12	10	89
images	848	248	123	1,219
share	69.6%	20.3%	10.1%	100%
1_overall	848	248	123	1,219
2_handwritten	9,766	2,776	1,309	13,851
3_typography	6,398	1,563	1,301	9,262
4_illustration	775	229	115	1,119
5_stamp	254	72	43	369

It was chosen by random search over PID assignments (200,000 draws, seed 20260917), minimizing deviation from 70/20/10 in both image count and per-class box count, so that the rare 5_stamp class stays present in every split.

model.train(data="data/data_pid.yaml", epochs=100, imgsz=640)

Only 8 of the 123 PID-level test images were in the original test split, so the two splits are not comparable — numbers obtained with data_pid.yaml should not be placed next to numbers obtained with data.yaml.

Usage (🤗 datasets)

Each data/images/<split>/ directory carries a metadata.jsonl, so the annotations load together with the images:

from datasets import load_dataset

ds = load_dataset("nakamura196/ndl-layout-dataset")
ds["train"][0]["objects"]
# {'bbox': [[x, y, w, h], ...], 'categories': [1, 1, 0, ...]}

# PID レベルの分割で取り出す
train = ds["train"].filter(lambda r: r["pid_split"] == "train")

Boxes in metadata.jsonl are in COCO pixel format [x_min, y_min, width, height] — not the normalized cx cy w h of the .txt label files. Columns: objects (bbox, categories), category_names, pid, page, width, height, num_objects, pid_split. Note that the train / val / test keys of the loaded dataset follow the original split (the val directory loads as validation); pid_split carries the PID-level one.

Dataset Creation

Source Data

NDL-DocL is built from public-domain materials in the NDL Digital Collections. Only out-of-copyright materials are included. The layout annotations were newly created by NDL staff and released in Pascal VOC XML.

Processing applied here

1. The distributed Pascal VOC XML annotations were converted to YOLO normalized cx cy w h text files.
2. The distributed JPEG images were downscaled so that the longer side is 640 px (a handful are 639 px). 1,141 of the 1,219 images are 640×480 — exactly 1/2.5 of the 1,600×1,200 images in the source archive — and 50 distinct sizes occur in total. Full-resolution images can be retrieved from NDL Digital Collections via the PID in the file name.
3. Random 70 / 20 / 10 split into train / validation / test (853 / 244 / 122), assigned per page, not per item (see Limitations).
4. Added in September 2026, without touching the image or label files: a PID-level split (data_pid.yaml, pid_*.txt) and metadata.jsonl files exposing the boxes to 🤗 datasets. Both are reproducible from scripts/. The script used for steps 1–3 in 2024 was not kept and no longer exists.

Bias, Risks, and Limitations

• Split leakage. The split is random per page, not per item. The 1,219 pages come from only 89 distinct PIDs, and pages from the same item appear across splits (34 of the 71 training PIDs also occur in validation, 32 in test). Because pages of the same book share typeface, layout and ownership seals, test scores from this split overestimate generalization to unseen materials. For a realistic estimate, re-split by PID (e.g. GroupShuffleSplit grouped on the PID prefix).
• Class imbalance. 5_stamp (369 boxes) and 4_illustration (1,119 boxes) are rare compared with text lines; per-class metrics should be reported, not only mAP.
• Resolution loss. Images are downscaled to 640 px on the long side, so thin text lines and small seals carry less detail than in the source data.
• Domain. Pre-modern Japanese materials only (woodblock-printed books, Chinese classics, ukiyo-e). Models trained here will not transfer to post-Meiji printed books, which NDL annotates with a different label scheme.
• Scale. 89 items is a small number of distinct sources for a layout model.

License and terms of use

The source NDL-DocL dataset is published under the Public Domain Mark (PDM) https://creativecommons.org/publicdomain/mark/1.0/. This derived dataset is redistributed on the same terms.

NDL asks users of NDL-DocL to observe the following (these are requests, not a legal contract). 国立国会図書館は NDL-DocL の二次利用にあたり、以下への配慮を求めています。

• データを編集・加工等して利用する場合は、それを行ったことを記載してください。 編集・加工等を、元となる作品・原資料の作者や当館が行なったかのような態様で公表 しないようご留意ください。
• 当該データが自由に二次利用可能であることの表記を保持してください。
• 元となる作品や、その作者の名声を傷つける形での利用は行わないようご留意ください。 また、元となる作品に関わる文化やコミュニティへの配慮を行ってください。
• 著作権以外の権利（著作者人格権、著作隣接権、肖像権、パブリシティ権、プライバシー権、 商標権等）にも留意し、関連法令を遵守してください。
• 論文等に利用する場合には、先行研究や後続研究と比較を容易にするため NDL-DocL データセットとバージョンの明記にご協力ください。

Please cite the NDL-DocL dataset with its version in publications, as requested above.

Citation

Source dataset:

@misc{ndl_docl,
  title        = {NDL-DocL Dataset (資料画像レイアウトデータセット), version 1.0},
  author       = {{National Diet Library, Japan}},
  year         = {2019},
  howpublished = {\url{https://github.com/ndl-lab/layout-dataset}}
}

This YOLO-formatted derivative:

@misc{nakamura_ndl_layout_yolo,
  title        = {NDL-DocL Kotenseki Layout Dataset (YOLO format)},
  author       = {Nakamura, Satoru},
  howpublished = {\url{https://huggingface.co/datasets/nakamura196/ndl-layout-dataset}}
}

Dataset Card Contact

Satoru Nakamura — https://huggingface.co/nakamura196