Question about Covariate shifted vs Near-OOD vs Far-OOD datasets

[dhir12314]

Hello!
I am currently working through the OpenOOD v1.5 paper and wondering how the datasets were classified as Near-OOD vs Far-OOD vs Covariate shifted. Were there any quantitative metrics or statistical techniques like PCA used to classify these datasets as such?

For instance, in the OpenOOD v1.5 paper how was NINCO determined to be a Near-OOD dataset. By visual inspection, it seems that it has a “greater” semantic shift and a “lesser” covariate shift when compared to ImageNet. Were there any metrics used to determine the threshold for “greater” or “lesser” in this case?

Thanks for your time and consideration!