How the pixel-level performance for MNIST, CIFAR-10, and ImageNet, etc are calculated?

[willyfh]

From what I understand, the original MNIST, CIFAR-10, and ImageNet do not provide the ground truth for the mask. But in Table 4 of your paper, we can see that the pixel-level performance for AUROC and PRO are reported.

How were you able to calculate the pixel-level performance for AUROC and PRO? Could you elaborate or point out in your paper how the pixel-level performances were computed on those datasets? Did you consider all pixels in the "abnormal images" as anomalous for the masks?