The M6Doc dataset for the research of document layout analysis in Modern Document is released by the Deep Learning and Visual Computing Lab of South China University of Technology.
💡 Notice: Our test data is completely free and openly accessible! You can directly download
M6Doc_test.zipwithout any application or password required.
| Dataset Split | Platform | Download Link | Format / Size | Password Required / Status |
|---|---|---|---|---|
| Full Dataset | Hugging Face | 🤗 hiuyi/M6Doc Repository | M6Doc.zip |
Yes (encrypted archive) |
| Full Dataset | Baidu Cloud | Download via BaiduNetdisk | 12.45 GB (Extract Code: xx3k) |
Yes (decompression password) |
| Test Data | Hugging Face | 🤗 Download M6Doc_test.zip | M6Doc_test.zip |
No Password |
You can download M6Doc.zip or the free M6Doc_test.zip using any of the following methods:
Method 1: Direct Web Download
- Go directly to the repository file list: hiuyi/M6Doc / Files and versions and click download on
M6Doc.ziporM6Doc_test.zip.
Method 2: Using Python (huggingface_hub)
from huggingface_hub import hf_hub_download
# Download Test Data (Free & Open Access)
hf_hub_download(
repo_id="hiuyi/M6Doc",
filename="M6Doc_test.zip",
repo_type="dataset",
local_dir="./"
)
# Download Full Dataset (Password Required)
hf_hub_download(
repo_id="hiuyi/M6Doc",
filename="M6Doc.zip",
repo_type="dataset",
local_dir="./"
)Method 3: Using Hugging Face CLI
# Download Test Data directly
huggingface-cli download --repo-type dataset hiuyi/M6Doc M6Doc_test.zip --local-dir ./
# Download Full Dataset
huggingface-cli download --repo-type dataset hiuyi/M6Doc M6Doc.zip --local-dir ./The full M6Doc dataset can only be used for non-commercial research purposes. The dataset is publicly accessible but encrypted with an additional password (except for the test data, which is completely free). To request access for the full dataset, please follow these steps:
Have this document signed and stamped by your institution. Please also prepare 1–2 recent publications (within the last 6 years) as evidence that you or your team conduct research in OCR, handwriting analysis and recognition, document image processing, or visual information extraction.
Upload both signed documents through the portal and fill out the "Recent Publications" block. Your application will be reviewed manually and you will be notified by email once a decision has been made (typically within 1–5 business days).
After approval, you will receive the decompression password via email.
⚠️ All users must comply with the use conditions at all times; failure to do so will result in revocation of access.
The M6Doc dataset should be used and distributed under the Creative Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0) License for non-commercial research purposes.
The M6Doc dataset contains a total of 9,080 modern document images, which are categorized into seven subsets, i.e., scientific article (11%), textbook (23%), test paper (22%), magazine (22%), newspaper (11%), note (5.5%), and book (5.5%) according to their content and layouts. It contains three formats: PDF (64%), photographed documents (5%), and scanned documents (31%). The dataset includes a total of 237,116 annotated instances.
The M6Doc datasets were collected from various sources, including arXiv, the official website of the Chinese People's Daily, and VKontakte. The source and composition of different subsets are shown below:
- The scientific article subset includes articles obtained by searching with the keywords "Optical Character Recognition" and "Document Layout Analysis" on arXiv. PDF files were then downloaded and converted to images.
- The textbook subset contains 2,080 scanned document images from textbooks for three grades (elementary, middle, and high school) and nine subjects (Chinese, Math, English, Physics, Chemistry, Biology, History, Geography, and Politics).
- The test paper subset consists of 2,000 examination papers covering the same nine subjects as the textbook subset.
- The magazine subset includes 1,000 Chinese and English magazines in PDF format, respectively. The Chinese magazines were sourced from five publishers: Global Science, The Mystery, Youth Digest, China National Geographic, and The Reader. The English magazines were sourced from five American publishers: The New Yorker, New Scientist, Scientific American, The Economist, and Time USA.
- The newspaper subset contains 500 PDF document images from the Chinese People's Daily and the Wall Street Journal.
- The note subset consists of students' handwritten notes in nine subjects, including 500 scanned pages.
- The book subset contains 500 photographed images, which were acquired from 50 books with 10 pages each. Each book has a distinct layout, resulting in considerable diversity in this subset.
To ensure that the definition of document layout elements is reasonable and traceable, we reviewed relevant information, such as layout knowledge and layout design. We also used knowledge from the book "Page Design: New Layout & Editorial Design (2019)" and referred to layout guidelines. In most cases, we followed the Wikipedia definition. Consequently, we defined 74 detailed document annotation labels.
The key factors in selecting these annotation labels include (1) the commonality of annotation labels between different document types, (2) the specificity of labels between different document types, (3) the frequency of labels, and (4) the recognition of independent pages. Figure 1 shows annotation samples of M6Doc. There are a total of 74 annotation categories in our dataset.
Figure 1. Example annotations of the M6Doc. Zoom in for better view.
Table 2 summarizes the overall frequency and distribution of labels.
Table 2. M6Doc dataset overview.
We provide a detailed annotation guideline (guideline_chinese.pdf, over 170 pages) and some typical annotation examples. 47 annotators performed the annotation task strictly according to the guidelines.
Once M6Doc.zip is decompressed, the dataset is organized in the following format:
├── M6Doc
├── annotations
│ ├── instances_train2017.json
│ └── instances_val2017.json
├── train2017
│ ├── xxx.jpg
│ └── ...
└── val2017
├── xxx.jpg
└── ...
Please consider to cite our paper when you use our dataset:
@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}
}While this work primarily focuses on modern documents, we are also conducting research on the layout analysis of Ancient Chinese Books. Please refer to the SCUT-CAB Dataset Release and guideline_Ancient for more details.
For any questions about the dataset, please contact the authors by sending an email to Prof. Jin (eelwjin@scut.edu.cn, or lianwen.jin@gmail.com).