As far as I understand it; Jev et al can be considered Zero Shot Classifiers.
It’s likely the system you built was trained on images of just cats and dogs; but Jev is using a general training set, and there is tuned for general purpose zero shot classification, rather than domain specific “is this a cat or a dog”.
It’s likely the system you built was trained on images of just cats and dogs; but Jev is using a general training set, and there is tuned for general purpose zero shot classification, rather than domain specific “is this a cat or a dog”.