pdf-layout-chinese Dataset Overview

Basic Information

• Name: pdf-layout-chinese
• License: AFL-3.0
• Task Type: Feature Extraction
• Languages: English, Chinese
• Size Category: 100M < n < 1B

Content

• Description: pdf-layout-chinese is a Chinese document layout analysis dataset targeting scholarly paper scenarios.
• Labels: 10 classes – text, title, image, image title, table, table title, header, footer, caption, formula.
• Splits: 5,000 training images and 1,000 validation images, stored in train and val directories respectively.
• Annotation Files: Each image has a same‑named JSON annotation file generated with labelme.

Annotation Format

• Tool: labelme
• Structure: Aligns with labelme format and includes key fields:
  • shapes: list of annotation instances.
  • labels: class labels.
  • points: polygon coordinates.
  • shape_type: polygon
  • imagePath: image path/name
  • imageHeight: image height
  • imageWidth: image width

Conversion

• Conversion Tool: labelme2coco.py
• Commands:
  • Training set: python3 labelme2coco.py train train_save_path --labels labels.txt
  • Validation set: python3 labelme2coco.py val val_save_path --labels labels.txt
• Output Location: Saved under train_save_path / val_save_path directories.