Hi Herbert,

Zero-shot learning is relatively a new topic of academic research. Although it has great potential and can be considered as ultimate goal of the object recognition task, result accuracies of the method are still low.

It is good to have in mind that such studies are around, however I don’t recommend it yet to use in a project that has practical expectation. I suppose your work as an intern is finding a relatively simple solution with high accuracy. You can try to find a way to increase the number of data and its variety.

Bon travail.