WHOOPS!
收藏资源简介:
WHOOPS!数据集是由希伯来大学和本古里安大学合作创建的,包含500张由设计师使用Midjourney等公开图像生成工具创建的合成图像。这些图像故意违反常识,旨在挑战AI模型对常识和组合性的理解。数据集不仅包含图像,还有详细的描述和解释,以及视觉问答对,用于评估模型在解释生成、图像描述、跨模态匹配和视觉问答等任务上的表现。WHOOPS!数据集的应用领域主要集中在提升AI模型的视觉常识推理能力,特别是在处理非典型、不合逻辑的图像场景时。
The WHOOPS! Dataset was collaboratively developed by the Hebrew University of Jerusalem and Ben-Gurion University of the Negev. It comprises 500 synthetic images created by designers using public image generation tools such as Midjourney. These images are deliberately designed to violate common sense, aiming to challenge AI models' understanding of common sense and compositionality. In addition to the image samples, the dataset also includes detailed descriptions and explanations, as well as visual question-answering pairs, which are used to evaluate model performance on tasks including explanation generation, image captioning, cross-modal matching and visual question answering. The main application areas of the WHOOPS! Dataset focus on enhancing the visual common-sense reasoning capabilities of AI models, especially when dealing with atypical and illogical image scenes.

- 1Breaking Common Sense: WHOOPS! A Vision-and-Language Benchmark of Synthetic and Compositional Images希伯来大学 · 2023年



