kbang2021/doclaynet-6class

DocLayNet 6-Class Filtered Dataset

Dataset Description

This is a filtered version of the DocLayNet dataset containing only 6 most relevant layout element classes for document layout analysis tasks.

Original Dataset

DocLayNet is a human-annotated document layout segmentation dataset containing 80,863 pages from diverse sources with 11 distinct layout categories.

Citation:

@article{doclaynet2022,
  title = {DocLayNet: A Large Human-Annotated Dataset for Document-Layout Analysis},
  author = {Pfitzmann, Birgit and Auer, Christoph and Dolfi, Michele and Nassar, Ahmed S and Staar, Peter W J},
  year = {2022},
  doi = {10.1145/3534678.3539043},
}

Filtering Methodology

Classes Retained (6):

1. Text - Body text paragraphs
2. List-item - List elements (bulleted, numbered)
3. Section-header - Section and subsection titles
4. Picture - Images, figures, diagrams
5. Table - Tabular data structures
6. Caption - Image and table captions

Classes Removed (5):

• Footnote
• Formula
• Page-footer
• Page-header
• Title

Rationale: Focus on the most common and semantically important layout elements for general document understanding tasks. The 6 retained classes represent 85.1% of all annotations in the original dataset.

Dataset Statistics

Split Distribution

Split	Images	Annotations	Classes
Train	68,673	800,614	6
Validation	6,446	85,057	6
Test	4,952	56,483	6
Total	80,071	942,154	6

Class Distribution (Training Set)

Based on 800,614 annotations:

Class ID	Class Name	Count	Percentage
0	Caption	19,218	2.4%
1	List-item	161,818	20.2%
2	Picture	39,667	5.0%
3	Section-header	118,590	14.8%
4	Table	30,070	3.8%
5	Text	431,251	53.9%

Retention from Original Dataset

• Images retained: 99.0%
• Annotations retained: 85.1%

Dataset Structure

Format

Annotations are provided in COCO JSON format:

DocLayNet_6class/
├── coco/
│   ├── train.json      # Training annotations
│   ├── val.json        # Validation annotations
│   └── test.json       # Test annotations
└── README.md           # This file

Images are NOT included - use the original DocLayNet image files from:

• HuggingFace: docling-project/DocLayNet
• Official source: https://github.com/DS4SD/DocLayNet

Loading the Dataset

Using HuggingFace Datasets

from datasets import load_dataset

# Load the filtered annotations
dataset = load_dataset("kbang2021/doclaynet-6class")

# Access splits
train_data = dataset["train"]
val_data = dataset["validation"]
test_data = dataset["test"]

Manual Loading

import json
from pathlib import Path

# Load COCO annotations
with open("coco/train.json") as f:
    train_coco = json.load(f)

# Categories
categories = train_coco["categories"]  # 6 classes with IDs 0-5

# Images
images = train_coco["images"]  # Image metadata

# Annotations
annotations = train_coco["annotations"]  # Bounding boxes

Annotation Format

Each annotation follows the COCO format:

{
  "id": 12345,
  "image_id": 123,
  "category_id": 5,  // 0-5 (remapped from original 11 classes)
  "bbox": [x_min, y_min, width, height],  // In pixels
  "area": 12345.67,
  "iscrowd": 0
}

Category Mapping

Original DocLayNet → 6-Class Filtered:

Original ID	Original Name	Filtered ID	Filtered Name	Status
0	Caption	0	Caption	✅ Kept
1	Footnote	-	-	❌ Removed
2	Formula	-	-	❌ Removed
3	List-item	1	List-item	✅ Kept
4	Page-footer	-	-	❌ Removed
5	Page-header	-	-	❌ Removed
6	Picture	2	Picture	✅ Kept
7	Section-header	3	Section-header	✅ Kept
8	Table	4	Table	✅ Kept
9	Text	5	Text	✅ Kept
10	Title	-	-	❌ Removed

Use Cases

This filtered dataset is ideal for:

• Document layout analysis with focus on content structure
• Information extraction from documents (text, tables, figures)
• Object detection model training for document AI
• Multi-scale document understanding tasks
• Transfer learning from general object detection to document analysis

Limitations

1. Images not included: You must obtain images from the original DocLayNet dataset
2. Class imbalance: Text class dominates (53.9% of annotations)
3. Domain specific: Focused on document layout, may not generalize to other domains
4. Annotation quality: Inherits any annotation errors from original DocLayNet

Ethical Considerations

• Dataset maintains the original DocLayNet license (CDLA-Permissive-2.0)
• No personal or sensitive information in annotations
• Source documents from diverse domains (financial, scientific, patents, manuals)
• Should not be used to discriminate based on document type or origin

Citation

If you use this filtered dataset, please cite both:

1. Original DocLayNet paper:

@article{doclaynet2022,
  title = {DocLayNet: A Large Human-Annotated Dataset for Document-Layout Analysis},
  author = {Pfitzmann, Birgit and Auer, Christoph and Dolfi, Michele and Nassar, Ahmed S and Staar, Peter W J},
  year = {2022},
  doi = {10.1145/3534678.3539043},
}

2. This filtered version:

@misc{doclaynet6class2024,
  title = {DocLayNet 6-Class: Filtered Document Layout Analysis Dataset},
  author = {[Keng Boon, Ang]},
  year = {2026},
  howpublished = {\url{https://huggingface.co/datasets/kbang2021/doclaynet-6class}},
  note = {Filtered subset of DocLayNet containing 6 primary layout element classes}
}

License

This filtered dataset maintains the original license:

CDLA-Permissive-2.0 (Community Data License Agreement – Permissive – Version 2.0)

See: https://cdla.dev/permissive-2-0/

Acknowledgments

• Original DocLayNet dataset: IBM Research
• Built using the layout-for-tools evaluation framework

Contact

For questions or issues with this filtered dataset, please open an issue on the repository.

For questions about the original DocLayNet dataset, see: https://github.com/DS4SD/DocLayNet