They work adequately for checking to see if a person claiming to be John Q. has roughly the right facial features and skin tone- because you are just making a s

They work adequately for checking to see if a person claiming to be John Q. has roughly the right facial features and skin tone- because you are just making a single comparison. It doesn’t work well in the other direction- finding which DL photo might belong to an unknown John Doe. Given the age of many DL photos (mine is from 8 years ago) and the fact that they are from a single aspect angle, they are not suited for the reverse task of finding which DL image is a good enough match for a person’s face from surveillance footage. If you tried to search every state database for a white guy with brown hair, and account for the fact that he may have a different beard or hair style from the DL photo that may be years old, you’re not going to narrow the pool enough to help the FBI. If you made your match threshold low enough to be reasonably reasonably confident that you didn’t miss the guy, you’d end up with millions of matches. The only decent facial recognition techniques use more reference images and use other methods to narrow the search space. The reason that Facebook was able to get their automated photo tagging working years ago was because they had exquisite knowledge of your social graph and only matched people 1-2 connections away from you. But a nationwide search for an unknown criminal doesn’t have the ability to drastically narrow the search space like this.