Visual Question Answering with Text in Images

The TextVQA dataset contains 45,336 questions based on 28,408 images from the OpenImages collection. It requires models to read and reason about text present within images to answer the provided questions.

Use Cases

• Train a multimodal model to answer questions by reading text present in 28,408 images.
• Benchmark model performance on the visual question answering task using 45,336 question-answer pairs.
• Develop reasoning capabilities that integrate visual content with textual information extracted from images.

Strengths

• Contains 45,336 question-answer pairs, providing substantial data for model training.
• Based on 28,408 images from the established OpenImages dataset.
• Specifically designed for the challenging task of reasoning about text within images.

Limitations

• Annotations are crowdsourced, which may introduce label noise or inconsistency.
• The dataset's geographic coverage is indicated as US, which may limit generalizability to other regions.
• The dataset is monolingual (English), restricting its use for multilingual model development.

Provenance

Source

Facebook, derived from the OpenImages dataset.

Collection Method

Crowdsourced annotations.

Freshness

Last updated on 2024-01-18.

Geography

Primarily United States (US).