ImageNet-E

ImageNet-E是由阿里巴巴集团创建的一个数据集，专注于通过编辑物体属性如背景、大小、位置和方向来评估图像分类器的鲁棒性。该数据集包含47872张图片，覆盖373个类别，主要来源于ImageNet。创建过程中，使用了基于扩散模型的对象编辑工具，自动生成具有控制属性的图像。ImageNet-E的应用领域主要集中在提高计算机视觉模型的鲁棒性，特别是在对抗性攻击和常见图像损坏的场景中。

ImageNet-E is a dataset created by Alibaba Group, which focuses on evaluating the robustness of image classifiers by editing object attributes such as background, size, position and orientation. This dataset contains 47,872 images covering 373 categories, and is mainly derived from ImageNet. During its creation, diffusion model-based object editing tools were employed to automatically generate images with controllable attributes. The application of ImageNet-E is primarily focused on improving the robustness of computer vision models, especially in scenarios of adversarial attacks and common image corruptions.