Voxel51/ImageNet-O

ImageNet-O数据集包含与标准ImageNet-1k数据集中不同的类别的图像，用于测试在ImageNet-1k上训练的计算机视觉模型的鲁棒性和分布外检测能力。该数据集包含的类别与ImageNet-1k中的1000个类别不同，用于评估模型在处理分布外样本时的性能。数据集由Dan Hendrycks等人策划，使用MIT许可证发布，并提供了相关的GitHub仓库和论文链接。

The ImageNet-O dataset consists of 2000 samples from classes not found in the standard ImageNet-1k dataset. It is designed to test the robustness and out-of-distribution detection capabilities of computer vision models trained on ImageNet-1k. The dataset contains images from classes distinct from the 1,000 classes in ImageNet-1k, enabling the testing of model performance on out-of-distribution samples, i.e., images that are semantically different from the training data. It is commonly used to evaluate out-of-distribution detection methods for models trained on ImageNet, using the Area Under the Precision-Recall curve (AUPR) metric. The dataset is manually annotated, has a naturally diverse class distribution, and is large scale. It is curated by Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Steinhardt, and Dawn Song, and shared by Harpreet Sahota, Hacker-in-Residence at Voxel51. The dataset is in English and is licensed under the MIT License.