DocLayNet is a human-annotated document layout segmentation dataset containing 80863 pages from a broad variety of document sources.
| Jan. 26th, 2023 | The DocLayNet dataset is available on Hugging Face ds4sd/DocLayNet. |
| Jan. 13th, 2023 | We are hosting a competition on layout segmentation in corporate documents in ICDAR 2023. Find the details in the competition website. |
DocLayNet provides page-by-page layout segmentation ground-truth using bounding-boxes for 11 distinct class labels on 80863 unique pages from 6 document categories. It provides several unique features compared to related work such as PubLayNet or DocBank:
- Human Annotation: DocLayNet is hand-annotated by well-trained experts, providing a gold-standard in layout segmentation through human recognition and interpretation of each page layout
- Large layout variability: DocLayNet includes diverse and complex layouts from a large variety of public sources in Finance, Science, Patents, Tenders, Law texts and Manuals
- Detailed label set: DocLayNet defines 11 class labels to distinguish layout features in high detail.
- Redundant annotations: A fraction of the pages in DocLayNet are double- or triple-annotated, allowing to estimate annotation uncertainty and an upper-bound of achievable prediction accuracy with ML models
- Pre-defined train- test- and validation-sets: DocLayNet provides fixed sets for each to ensure proportional representation of the class-labels and avoid leakage of unique layout styles across the sets.
The DocLayNet dataset is available on Hugging Face at ds4sd/DocLayNet.
>>> from datasets import load_dataset
>>> dataset = load_dataset("ds4sd/DocLayNet")
>>> dataset
DatasetDict({
train: Dataset({
features: ['image_id', 'image', 'width', 'height', 'doc_category', 'collection', 'doc_name', 'page_no', 'objects'],
num_rows: 69375
})
validation: Dataset({
features: ['image_id', 'image', 'width', 'height', 'doc_category', 'collection', 'doc_name', 'page_no', 'objects'],
num_rows: 6489
})
test: Dataset({
features: ['image_id', 'image', 'width', 'height', 'doc_category', 'collection', 'doc_name', 'page_no', 'objects'],
num_rows: 4999
})
})| Dataset | Records | Size(GB) | URL |
|---|---|---|---|
| DocLayNet core dataset | 80,863 | 28 GiB | Download |
| DocLayNet extra files | 80,863 | 7.5 GiB | Download |
Additionally, we provide the labeling guideline used for training of the annotation experts here.
DocLayNet provides four types of data assets:
- PNG images of all pages, resized to square
1025 x 1025px - Bounding-box annotations in COCO format for each PNG image
- Extra: Single-page PDF files matching each PNG image
- Extra: JSON file matching each PDF page, which provides the digital text cells with coordinates and content
The dataset is organized in the following directory structure:
Doclaynet core dataset
├── COCO
│ ├── test.json
│ ├── train.json
│ └── val.json
├── PNG
│ ├── <hash>.png
│ ├── ...
Doclaynet extra files
├── PDF
│ ├── <hash>.pdf
│ ├── ...
├── JSON
│ ├── <hash>.json
│ ├── ...
Document pages are provided as bitmap images (PNG) and in original PDF format. The layout annotations in COCO format are referring to the PNG images only.
For bounding-box annotations, DocLayNet provides standard COCO format annotations as defined here. Each COCO image record contains additional custom fields to allow data sub-selection and provide provenance.
...
{
"id": 1,
"width": 1025,
"height": 1025,
"file_name": "132a855ee8b23533d8ae69af0049c038171a06ddfcac892c3c6d7e6b4091c642.png",
// Custom fields:
"doc_category": "financial_reports" // high-level document category
"collection": "ann_reports_00_04_fancy", // sub-collection name
"doc_name": "NASDAQ_FFIN_2002.pdf", // original document filename
"page_no": 9, // page number in original document
"precedence": 0, // Annotation order, non-zero in case of redundant double- or triple-annotation
},
...The doc_category field uses one of the following constants:
financial_reports,
scientific_articles,
laws_and_regulations,
government_tenders,
manuals,
patents
The annotation records shown below all belong to the page image shown above. Every bounding-box is a separate record, matched to the common image_id.
Click to expand...
"annotations": [
{
"id": 8,
"image_id": 1,
"category_id": 1,
"bbox": [
210.06018382352943,
31.14536268939389,
173.9850743464052,
39.270946654040586
],
"segmentation": [
[
210.06018382352943,
31.14536268939389,
210.06018382352943,
70.41630934343448,
384.04525816993464,
70.41630934343448,
384.04525816993464,
31.14536268939389
]
],
"area": 6832.558573256964,
"iscrowd": 0,
"precedence": 0
},
{
"id": 9,
"image_id": 1,
"category_id": 7,
"bbox": [
434.9334063800317,
-0.4906348977078778,
589.8585905504372,
590.2337819428021
],
"segmentation": [
[
434.9334063800317,
-0.4906348977078778,
434.9334063800317,
589.7431470450942,
1024.791996930469,
589.7431470450942,
1024.791996930469,
-0.4906348977078778
]
],
"area": 348154.46671203536,
"iscrowd": 0,
"precedence": 0
},
{
"id": 10,
"image_id": 1,
"category_id": 8,
"bbox": [
66.99346405228758,
112.10344760101009,
290.869358251634,
13.66279703282828
],
"segmentation": [
[
66.99346405228758,
112.10344760101009,
66.99346405228758,
125.76624463383837,
357.8628223039216,
125.76624463383837,
357.8628223039216,
112.10344760101009
]
],
"area": 3974.0890048610913,
"iscrowd": 0,
"precedence": 0
},
{
"id": 11,
"image_id": 1,
"category_id": 10,
"bbox": [
66.99346405228758,
133.5865287247475,
325.3560694444444,
131.31064046717177
],
"segmentation": [
[
66.99346405228758,
133.5865287247475,
66.99346405228758,
264.89716919191926,
392.34953349673196,
264.89716919191926,
392.34953349673196,
133.5865287247475
]
],
"area": 42722.71385863161,
"iscrowd": 0,
"precedence": 0
},
{
"id": 12,
"image_id": 1,
"category_id": 10,
"bbox": [
66.99346405228758,
272.84557828282834,
325.4857017973857,
131.3025
],
"segmentation": [
[
66.99346405228758,
272.84557828282834,
66.99346405228758,
404.14807828282835,
392.4791658496732,
404.14807828282835,
392.4791658496732,
272.84557828282834
]
],
"area": 42737.08636025123,
"iscrowd": 0,
"precedence": 0
},
{
"id": 13,
"image_id": 1,
"category_id": 10,
"bbox": [
66.99346405228758,
414.8919160353536,
325.69678145424837,
80.83059406565656
],
"segmentation": [
[
66.99346405228758,
414.8919160353536,
66.99346405228758,
495.72251010101013,
392.6902455065359,
495.72251010101013,
392.6902455065359,
414.8919160353536
]
],
"area": 26326.26433021921,
"iscrowd": 0,
"precedence": 0
},
{
"id": 14,
"image_id": 1,
"category_id": 10,
"bbox": [
112.37566899509804,
626.7556887626263,
863.111772998366,
310.9754294507576
],
"segmentation": [
[
112.37566899509804,
626.7556887626263,
112.37566899509804,
937.7311182133839,
975.487441993464,
937.7311182133839,
975.487441993464,
626.7556887626263
]
],
"area": 268406.55427217163,
"iscrowd": 0,
"precedence": 0
},
{
"id": 15,
"image_id": 1,
"category_id": 5,
"bbox": [
18.32874183006536,
1005.2406660353536,
7.4496732026143775,
10.353535094696968
],
"segmentation": [
[
18.32874183006536,
1005.2406660353536,
18.32874183006536,
1015.5942011300506,
25.77841503267974,
1015.5942011300506,
25.77841503267974,
1005.2406660353536
]
],
"area": 77.13045294729152,
"iscrowd": 0,
"precedence": 0
},
...
]
DocLayNet provides auxiliary JSON files which contain the bounding-box coordinates and text for every PDF cell, and additional metadata. This data is generated only from the PDFs and is independent from the annotations.
The snippet below shows part of the JSON data for the page shown further above. The text cell shown is the section heading (index: 3).
{
"metadata": {
"page_hash": "132a855ee8b23533d8ae69af0049c038171a06ddfcac892c3c6d7e6b4091c642", // unique identifier, equal to filename
"original_filename": "NASDAQ_FFIN_2002.pdf", // original document filename
"page_no": 9, // page number in original document
"num_pages": 28, // total pages in original document
"original_width": 612, // width in pixels @72 ppi
"original_height": 792, // height in pixels @72 ppi
"coco_width": 1025, // with in pixels in PNG and COCO format
"coco_height": 1025, // with in pixels in PNG and COCO format
"collection": "ann_reports_00_04_fancy", // sub-collection name
"doc_category": "financial_reports" // high-level document category
},
"cells": [ // all text cells in the digital PDF data
{
// Bounding-box coordinates of text cells,
// formatted as [x,y,w,h] (same as COCO annotations)
// where (x,y) is the upper-left corner and
// (w,h) is the width and height
// in the coordinate space of (0,0, coco_width, coco_height)
"bbox": [
66.99346405228758,
112.10344760101009,
290.869358251634,
13.66279703282828
],
"text": "Leigh Taliaferro, M.D., values consistency.", // string content of cell
"font": {
"color": [
12,
72,
142,
255
],
"name": "/AAAAAC+HelveticaNeue-Medium",
"size": 1
}
},
..."DocLayNet: A Large Human-Annotated Dataset for Document-Layout Analysis" (KDD 2022).
- Birgit Pfitzmann (bpf@zurich.ibm.com)
- Christoph Auer (cau@zurich.ibm.com)
- Michele Dolfi (dol@zurich.ibm.com)
- Ahmed Nassar (ahn@zurich.ibm.com)
- Peter Staar (taa@zurich.ibm.com)
ArXiv link: https://arxiv.org/abs/2206.01062
Citation:
@article{doclaynet2022,
title = {DocLayNet: A Large Human-Annotated Dataset for Document-Layout Analysis},
doi = {10.1145/3534678.353904},
url = {https://arxiv.org/abs/2206.01062},
author = {Pfitzmann, Birgit and Auer, Christoph and Dolfi, Michele and Nassar, Ahmed S and Staar, Peter W J},
year = {2022}
}
