M⁶Doc: A Large-Scale Multi-Format, Multi-Type, Multi-Layout, Multi-Language, Multi-Annotation Category Dataset for Modern Document Layout Analysis

---

⚠️ Access Request & Application Instructions

The M⁶Doc dataset can only be used for non-commercial research purposes. The full dataset is publicly accessible but encrypted with an additional password (Notice: our test data is completely free and open-access / 免费开放, no application or password required). To request access for the full dataset, please follow these steps:

Step 1: Download and complete the agreement document:

• 📄 Application Form for Using M⁶Doc (Direct Download)

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.

Step 2: Submit your application online:

🔗 SCUT DLVC Lab Dataset Access Portal → Apply for M6Doc

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).

Step 3: Download the dataset:

After approval, you will receive the download link and 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.

---

Dataset Download

💡 Notice: Our test data is completely free and openly accessible (我们的 test 数据是免费开放的)! You can directly download M6Doc_test.zip without 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	免费开放 (No Password)

How to Download from Hugging Face

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.zip or M6Doc_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 ./

---

Dataset Overview

The M⁶Doc dataset contains a total of 9,080 modern document images, which are categorized into seven subsets:

• Scientific article (11%): Articles obtained by searching with the keywords "Optical Character Recognition" and "Document Layout Analysis" on arXiv.
• Textbook (23%): 2,080 scanned document images from textbooks for three grades (elementary, middle, and high school) and nine subjects.
• Test paper (22%): 2,000 examination papers covering the same nine subjects as the textbook subset.
• Magazine (22%): 1,000 Chinese magazines (Global Science, Youth Digest, China National Geographic, etc.) and 1,000 English magazines (The New Yorker, New Scientist, Scientific American, The Economist, Time, etc.).
• Newspaper (11%): 500 PDF document images from the Chinese People's Daily and the Wall Street Journal.
• Note (5.5%): Students' handwritten notes in nine subjects, including 500 scanned pages.
• Book (5.5%): 500 photographed images acquired from 50 books with 10 pages each.

Formats & Annotations

• Formats: PDF (64%), Scanned documents (31%), Photographed documents (5%).
• Instances: A total of 237,116 annotated instances across 74 detailed layout categories.

---

Data Annotation

Label Definition

To ensure that the definition of document layout elements is reasonable and traceable, relevant layout knowledge and layout design guidelines were reviewed. There are a total of 74 annotation categories in our dataset.

Figure 1. Example annotations of M⁶Doc. Zoom in for better view.

For a fair evaluation, the dataset is divided into training, validation, and test sets in a ratio of 6:1:3.

Table 2. M⁶Doc dataset overview.

Annotation Guideline

• 📖 guideline_chinese.pdf (over 170 pages)

---

Directory Format

Once M6Doc.zip is decompressed, the dataset is organized in the following standard COCO format:

├── M6Doc
    ├── annotations
    │   ├── instances_train2017.json
    │   ├── instances_val2017.json
    │   └── instances_test.json
    ├── train2017
    │   ├── xxx.jpg
    │   └── ...
    ├── val2017
    │   ├── xxx.jpg
    │   └── ...
    └── test2017
        ├── xxx.jpg
        └── ...

---

License

The M⁶Doc dataset should be used and distributed under the Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0) License for non-commercial research purposes.

---

Citation and Contact

Please cite our paper if you use this 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}
}

• Ancient Chinese Books Layout Analysis (Related Work): SCUT-CAB Dataset Release | Guideline Ancient (PDF)
• Contact: For inquiries, please contact Prof. Lianwen Jin at eelwjin@scut.edu.cn or lianwen.jin@gmail.com.