Anyone know how well these similarity neural nets deal with something as simple as, say, rotation? It would be trivial to generate 4 hashes for the 90° cardinal

Anyone know how well these similarity neural nets deal with something as simple as, say, rotation? It would be trivial to generate 4 hashes for the 90° cardinal rotations (and, I guess 4 more for axis flips), but how about, for example, 11°? It would really suck if everyone starts implementing this and there's such a simple way to circumvent it.