dek-receipt-cord

Annotated receipt dataset in CORD format for OCR-free document understanding.

Description

This dataset contains images of retail receipts annotated in the CORD (Clova OCR Dataset) format. The receipts were photographed in natural conditions, including realistic distortions such as perspective skew, varying lighting conditions, and partial occlusions.

The dataset is intended for research and development of Key Information Extraction (KIE) models, with a focus on OCR-free approaches such as Donut.

Contents

• 100 fully annotated receipt images
• 1000 raw receipt images collected (annotation currently in progress)

Only the fully annotated subset is included in the current dataset release.

Data Collection

• Images captured using mobile phone cameras
• Real-world conditions:
  • varying angles
  • uneven illumination
  • natural background noise

Annotation Format

• CORD format
• JSON annotations aligned with official CORD specification
• Key fields include:
  • store name
  • transaction date
  • purchased items (name, price)
  • subtotals and total amount

Intended Use

• Key Information Extraction (KIE)
• End-to-end document understanding
• Fine-tuning and evaluation of Donut-based models
• Benchmarking OCR-free approaches

Dataset Structure

dek-receipt-cord/ ├── images/ │ ├── 000001.jpg │ └── ... ├── annotations/ │ ├── 000001.json │ └── ... └── README.md

Limitations

• The dataset size is currently limited and is not sufficient for final metric-driven evaluation.
• Additional annotations are planned for future releases.

Related Resources

• Baseline model: https://huggingface.co/SvetaLana25/dek-receipt-donut-baseline