Wow. Figure 1 tells it all -- and highlights how unstable CNNs can be. I can't help but think that in a "real" environment a more traditional facial recognition

Wow. Figure 1 tells it all -- and highlights how unstable CNNs can be. I can't help but think that in a "real" environment a more traditional facial recognition algo would also be used simultaneously, for example measuring landmark distances, which would be likely to give a high probability of a match.