TextOCR Dataset for Scene Text Recognition

TextOCR provides over 1 million text annotations from real-world images. The dataset is designed for training and evaluating scene text recognition models, sourced from the TextVQA dataset. It contains word-level annotations on images from Open Images.

Use Cases

• Train word-level text recognition models using cropped image patches and their corresponding text annotations.
• Benchmark scene text detection performance using provided bounding box coordinates for each word.
• Improve multilingual OCR by leveraging annotations that include non-Latin scripts present in the source images.
• Develop end-to-end text understanding pipelines by combining the image data with the extracted textual content.

Strengths

• Over 1 million word annotations provide a substantial training corpus.
• Annotations include word-level bounding boxes and transcriptions for precise localization.

Limitations

• Dataset size and specific annotation counts are not detailed in the provided metadata.
• Potential class imbalance or geographic bias may exist based on the source image collection.

Provenance

Source

Derived from the TextVQA dataset, which uses images from Open Images.

Collection Method

Word-level annotations were manually added to images containing text.