Part of this problem is their using software L1. L1 has terrible performance but their EULA states that you can't refer to them by name in Academic papers. So their always refereed to anonymously as the"Commercial Competitor".
With that in their EULA, they are essentially stating: "We know that our software sucks, but to prevent others from stating that we specifically suck, we're going to forbid academic papers (which are, for the most part, impartial studies) from mentioning us by name, thereby putting doubt over whether it's us or some other facial recognition product."
I need to make a note to actually read EULAs from now on, because I want to get my money back if I ever buy software with that in the EULA.
My view might be slightly skewed since the basis of the competition was to make it hard (low light, noise, occlusion, pose, atmospheric blur) and the commercial competitors are trained on easier datasets (close range, well lit, uniform pose and expression). While L1's system would perform better on driver's license images false positives are going to occur in any system. So when it flags a person it should be regarded as something fishy might be going on instead the oracle says this person is a terrorist who buys alcohol for teens, lock him away!