A-OKVQA

A-OKVQA是一个新的基于知识的视觉问答基准。A-OKVQA是OK-VQA的增强后继者，包含一系列25k问题，需要广泛的常识和世界知识来回答。A-OKVQA中的问题具有挑战性，概念多样，需要图像之外的知识，与现有的基于知识的视觉问答数据集相比，它们不能通过简单地查询知识库来回答。为了简化使用无限知识源的工作，训练集中的问题与提供事实和回答这些事实所需的推理片段的理性配对。
回答问题所需的知识类型包括 (但不限于):
常识-人类从日常经验中学到的关于世界的知识 (例如，许多甜甜圈在购物车中制成，这意味着它们是出售的，而不是个人消费的)。
视觉-视觉上表示的概念的知识 (例如，静音的颜色托盘与20世纪50年代相关联)。
知识库-从教科书，维基百科和其他文本来源获得的知识 (例如，热狗是在奥地利发明的)。
物理-有关世界物理的知识 (例如，阴影区域的温度低于其他区域)。

A-OKVQA is a novel knowledge-based visual question answering benchmark. As an enhanced successor to OK-VQA, it encompasses 25,000 questions that demand broad commonsense and world knowledge for correct answering. The questions in A-OKVQA are challenging, cover diverse concepts, and require knowledge beyond the visual content of images; unlike existing knowledge-based visual QA datasets, they cannot be resolved by simply querying a single knowledge base. To streamline work with unbounded knowledge sources, questions in the training split are paired with rationales that provide the factual details and reasoning steps necessary to answer them.

Types of knowledge required to answer these questions include (but are not limited to):
1. Commonsense knowledge: Knowledge about the world learned through everyday human experience, e.g., "Many donuts are produced in commercial shopping carts, meaning they are intended for sale rather than personal consumption."
2. Visual knowledge: Knowledge of visually represented concepts, e.g., "Muted color palettes are associated with the 1950s."
3. Knowledge base knowledge: Knowledge acquired from textbooks, Wikipedia, and other textual sources, e.g., "Hot dogs were invented in Austria."
4. Physical knowledge: Knowledge about the physical rules of the world, e.g., "Areas in shadow have lower temperatures than other regions."