The panopticon is coming and its accuracy will improve. Facial recognition by itself may not improve much, but when paired with other information it's going to skyrocket in accuracy, and I suspect it'll happen in China first if it hasn't already. That other information will include things like the possible identities of other people in the same vicinity that have known associations, locations pulled from recognized devices, other fuzzy identifications along travel routes, gait analysis, matching facial recognitions with clothing identification to improve tracking, etc. None of the individual parts are that difficult, they just result in a massive amount of fuzzy data to be mined and computers can be very good at that mining.

Another factor is that this may be challenging when starting from a facial recognition point in a crowd and trying to identify a person, but what if you're starting from a list of "the usual suspects" who've been thoroughly photographed, had gait analysis done and had their social networks mapped? What if you start from "do we see any sign that any of these 50k individuals are in the crowd?"