OCR stands for optical character recognition. TextOCR is a dataset for detecting and recognizing text in arbitrary scenes. TextOCR provides about 1 million high-quality word annotations for images in TextVQA, and can perform end-to-end reasoning on downstream tasks such as visual question answering or image description.

The dataset includes:

• 28,134 images from the TextVQA dataset
• 903,096 annotated scene text words
• On average, each image has 32 related words.