데이터셋 탐색

The ImageNet-A dataset consists of real-world, unmodified, and naturally occurring examples that are misclassified by ResNet models. Source: On Robustness and Transferability of Convolutional Neural Networks Image Source: https://github.com/hendrycks/natural-adv-examples

We build a new evaluation set by adding spotting words to the images of ImageNet 2012 evaluation sets. There are 1,000 categories in ImageNet. For each category c, we find its most confusing category c*and spot the category name to every evaluation image. This evaluation set is challenging for many CLIP models. For example, OpenAI CLIP B-16 got a top-1 accuracy of as low as 32%, which is much lower than the original ImageNet evaluation set.

Tiny ImageNet-A is a subset of the Tiny ImageNet test set consisting of 3,374 images comprising real-world, unmodified, and naturally occurring examples that are misclassified by ResNet-18. The sampling process of Tiny ImageNet-A roughly follows the concept of ImageNet-A introduced by Hendrycks et al. ("Natural Adversarial Examples"). For further information on the sampling process visit the original paper.