Receipt Images With OCR Text And Key Information Classes

WildReceipt contains 1,765 photos of receipts, each annotated with OCR text boxes and 25 key information classes. The dataset provides 50,000 text boxes with bounding boxes and text, created by Theivaprakasham for benchmarking key information extraction from documents.

Use Cases

• Train a key information extraction model to predict one of the 25 classes for each OCR text box.
• Benchmark OCR accuracy and layout analysis using the provided bounding boxes and text fields across 1,765 receipt images.
• Develop a structured data parser to locate and extract specific fields like total amount or date from the 50,000 annotated text boxes.

Strengths

• Dataset provides 50,000 annotated text boxes for model training and evaluation.
• Includes 25 distinct classes for key information extraction tasks.
• Contains 1,765 real-world receipt images, offering visual diversity.

Limitations

• Limited to 1,765 images, which is a small sample size for training large vision-language models.
• Geographic and format bias is likely present as the source and receipt types are unspecified.
• Class distribution and potential label imbalance across the 25 classes are unknown.

Provenance

Source

Hugging Face, created by Theivaprakasham.

Collection Method

Collection of receipt photos annotated with OCR text boxes and class labels.

Freshness

Last updated in June 2022.

Geography

Region is listed as 'us', but specific coverage is unknown.