It feels like there are two sides of this: either recognition is amazing, either is really really far. It seems that after it generates the caption, this need

It feels like there are two sides of this: either recognition is amazing, either is really really far. It seems that after it generates the caption, this needs to be fed to some semantic pipe, so that a plane sitting on a book would not make sense, and try further. After all, it really depends on the training data. If the picture of a train ticket was never seen by the NN, how could it answer correctly? How ever, it should try to reduce the answer to some more meaningfull info, for example instead of two giraffes near a tree, ideally would have said, it's a text and would attempt OCR.