Open cainsmile opened 1 year ago
So, the target is to train on the clean dataset, like ImageNet, and show robustness on the corrupt dataset, like ImageNet-C, right?
Hi, yes, that is the setting of zero-shot robustness. Beyond that, we also expand the scope to Segmentation and Detection with the same settings.
So, the target is to train on the clean dataset, like ImageNet, and show robustness on the corrupt dataset, like ImageNet-C, right?