← AI Terminology
VQA - Visual Question Answering
Visual question answering (VQA) is the task of answering natural-language questions about an image’s content.
It is a core benchmark family for vision-language systems.
It is a core benchmark family for vision-language systems.
Why It Matters in AI
VQA tests grounded understanding: counting, attributes, relations, OCR. It drove multimodal architectures and remains a standard eval alongside captioning and retrieval.
Key Points
| Aspect | Description |
|---|---|
| Input | Image + question → answer |
| Metric | Accuracy / soft VQA scores |
| Models | Modern VLMs zero/few-shot VQA |
| Related | VLMs, image captioning, multimodal |
| Datasets | VQAv2, GQA, TextVQA, OK-VQA |
| Challenges | Language priors, compositionality, reading text in images |
Simple Analogy
Showing someone a photo and quizzing them — “How many red cups are on the table?” — language tied to what is seen.
Common Usage Examples
- Evaluate VLMs on VQAv2
- TextVQA for chart/OCR questions
- Attention/grounding visualisations
- Avoid models that answer without looking
Summary
In short: VQA answers natural-language questions about images — a flagship task for measuring grounded multimodal understanding.