license: cc-by-nc-nd-4.0
task_categories:
- object-detection
- image-segmentation
language:
- en
- zh
tags:
- document-layout-analysis
- document-understanding
- coco-format
- ocr
pretty_name: M6Doc (Test Set)
size_categories:
- 1K<n<10K
splits:
- name: test
num_examples: 2724
M6Doc — Test Set
⚠️ This dataset currently contains the test split only. The training and validation splits require an approved application from the dataset authors (see Access below).
Dataset Summary
M6Doc is a large-scale, multi-format document layout analysis dataset presented at CVPR 2023 by the Deep Learning and Visual Computing Lab of South China University of Technology (SCUT).
It covers 7 document types, 3 formats, and 74 fine-grained annotation categories — making it one of the most comprehensive datasets for modern document layout analysis.
| Split | Images | Annotations | Status |
|---|---|---|---|
| test | 2,724 | ~71K | ✅ Available here |
| train | ~5,450 | ~142K | 🔒 Requires application |
| validation | ~910 | ~23K | 🔒 Requires application |
Dataset Features
Each example contains:
| Column | Type | Description |
|---|---|---|
image |
Image |
Document image (viewable in browser) |
image_id |
int64 |
COCO image ID |
file_name |
string |
Original filename |
width |
int32 |
Image width in pixels |
height |
int32 |
Image height in pixels |
bboxes |
List[List[float]] |
Bounding boxes [x, y, w, h] per instance |
areas |
List[float] |
Area of each annotated instance |
category_ids |
List[int] |
Category ID per instance (1–74) |
category_names |
List[str] |
Category label per instance |
iscrowd |
List[int] |
Crowd flag per instance |
Document Categories
The dataset spans 7 document types:
| Type | Proportion |
|---|---|
| Textbook | 23% |
| Test Paper | 22% |
| Magazine | 22% |
| Scientific Article | 11% |
| Newspaper | 11% |
| Note (handwritten) | 5.5% |
| Book (photographed) | 5.5% |
And 3 document formats: PDF (64%), Scanned (31%), Photographed (5%).
Annotation Categories
74 detailed layout element categories including (but not limited to):
text, title, figure, figure_caption, table, table_caption, header, footer, page_number, footnote, formula, reference, list, abstract, author, affiliation, equation, sidebar, advertisement, ...
Usage
from datasets import load_dataset
ds = load_dataset("tuandunghcmut/M6Doc")
print(ds)
# DatasetDict({ test: Dataset({features: [...], num_rows: 2724}) })
# View a sample
sample = ds["test"][0]
sample["image"] # PIL Image
sample["category_names"] # list of label strings
sample["bboxes"] # list of [x, y, w, h]
Access
Test set: Freely available here.
Train & Validation sets: Require submitting an application form to the original authors:
- Fill in the Application Form
- Email to lianwen.jin@gmail.com or eelwjin@scut.edu.cn
- Include 1–2 recent publications in OCR / document analysis
License
This dataset is released under CC BY-NC-ND 4.0 — non-commercial research use only.
Citation
@InProceedings{Cheng_2023_CVPR,
author = {Cheng, Hiuyi and Zhang, Peirong and Wu, Sihang and Zhang, Jiaxin and Zhu, Qiyuan and Xie, Zecheng and Li, Jing and Ding, Kai and Jin, Lianwen},
title = {M6Doc: A Large-Scale Multi-Format, Multi-Type, Multi-Layout, Multi-Language, Multi-Annotation Category Dataset for Modern Document Layout Analysis},
booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
month = {June},
year = {2023},
pages = {15138-15147}
}