Voxel51/IllusionAnimals

这是一个包含2000个样本的FiftyOne数据集，名为IllusionAnimals。该数据集由伊朗科技大学的研究人员创建，用于评估多模态模型在识别和解释视觉错觉方面的能力，尤其是基于动物的错觉。数据集包括训练集3300个样本和测试集1100个样本，共有10个类别，包括猫、狗、鸽子、蝴蝶、大象、马、鹿、蛇、鱼和公鸡，还有一个“无错觉”类别。图像分辨率为512x512像素。数据集通过使用SDXL-Lightning模型创建基础图像，利用ControlNet模型生成错觉，并通过人类验证者进行标注。

This is a FiftyOne dataset named IllusionAnimals with 2000 samples. Created by researchers from IUST (Iran University of Science and Technology), it is designed to evaluate the ability of multimodal models to detect and interpret visual illusions, particularly animal-based ones. The dataset includes a training set of 3300 samples and a test set of 1100 samples, across 10 classes such as cat, dog, pigeon, butterfly, elephant, horse, deer, snake, fish, and rooster, with an additional No illusion class. The images are at a resolution of 512x512 pixels. The dataset was created using SDXL-Lightning for base images, transformed into illusions with ControlNet, and annotated by human validators.