Dataset Card for ImageNet A This is a FiftyOne dataset with 7450 samples. The recipe notebook for creating this FiftyOne Dataset can be found here. Installation If you haven't already, install FiftyOne: Usage Dataset Details Dataset Description ImageNet A is a dataset of adversarially filtered images that reliably fool current ImageNet classifiers. It contains natural, unmodified real world examples that transfer to various unseen ImageNet models, demonstrating that these models share weaknesses with adversarially selected images. These images cause consistent classification mistakes across various models. To create ImageNet A, the authors first downloaded numerous images related to an ImageNet class. They then deleted the images that fixed ResNet 50 classifiers correctly predicted. With the remaining incorrectly classified images, the authors manually selected visually clear images. The resulting ImageNet A dataset has ~7,500 adversarially filtered images. The ImageNet A dataset enables testing image classification performance when the input data distribution shifts[1]. ImageNet A can be used to measure model robustness to distribution shift using challenging natural images. The i…
We use cookies for essential functionality and analytics. You can accept or reject analytics cookies.Cookie policy