IllusionVQA
收藏资源简介:
IllusionVQA是由孟加拉工程技术大学创建的一个挑战性视觉错觉数据集,旨在测试视觉语言模型(VLM)对复杂光学错觉的理解和定位能力。该数据集包含435个实例,涵盖12种不同的光学错觉类别,每个实例包括一张包含错觉的图像、一个问题和多个选择题选项。数据集通过从互联网收集并手动筛选高质量的光学错觉图像,确保了数据的多样性和挑战性。IllusionVQA的应用领域包括评估和提升VLM在处理视觉错觉和复杂场景理解方面的能力,特别是在机器人导航和交互等实际应用中。
IllusionVQA is a challenging visual illusion dataset developed by Bangladesh University of Engineering and Technology. It is designed to evaluate the comprehension and localization abilities of vision-language models (VLMs) regarding complex optical illusions. The dataset comprises 435 instances covering 12 distinct optical illusion categories, with each instance containing an illusion-embedded image, a corresponding question, and several multiple-choice options. To ensure data diversity and challenge, high-quality optical illusion images were collected from the Internet and manually screened. Application domains of IllusionVQA include assessing and improving the capabilities of VLMs in handling visual illusions and complex scene understanding, particularly in practical scenarios such as robotic navigation and interaction.




