Consolidated Receipt Dataset with Normalized Bounding Boxes

The CORD (Consolidated Receipt Dataset) contains receipt images with normalized bounding box annotations for text parsing. The dataset is crowdsourced, monolingual, and categorized for token classification tasks. The size is estimated between 1,000 and 10,000 samples.

Use Cases

• Train a token classification model on receipt text using normalized bounding box coordinates for entity recognition.
• Develop a parsing pipeline for structured data extraction from receipt images based on the provided annotations.
• Analyze the spatial layout of text elements on receipts using the normalized bounding box data for document understanding.

Strengths

• Dataset size is categorized as between 1,000 and 10,000 samples, providing a moderate scale for model training.
• Annotations are normalized bounding boxes, which provide consistent spatial coordinates for text elements.
• The dataset is specifically designed for the token classification task, offering clear task alignment.

Limitations

• The exact number of rows, columns, and file formats is unknown, limiting precise technical assessment.
• Data is crowdsourced, which may introduce label noise or inconsistency in bounding box annotations.
• The dataset's geographic focus is the US region, limiting generalizability to receipts from other countries.

Provenance

Source

Original source dataset, author wkrl.

Collection Method

Crowdsourced annotation creation.

Freshness

Last updated in July 2022.

Geography

US region.