VQA: Visual Question Answering

The paper introduces Visual Question Answering (VQA), a task in which a system receives an image and a free-form, open-ended natural language question about that image and must return an accurate natural language answer. Mirroring real-world uses such as assisting visually impaired people, the questions selectively probe different regions of an image, including background details and underlying context, so a successful system needs more detailed visual understanding and more complex reasoning than one that produces generic image captions.

To support research, the authors provide a large benchmark of roughly 0.25 million images, 0.76 million questions, and 10 million answers, noting that VQA is amenable to automatic evaluation because many answers are short or fit a multiple-choice format. They supply numerous baselines and methods and compare them against human performance, establishing VQA as a challenging benchmark that bridges computer vision and natural language processing.