It’s got a minimum 80% error rate. More work to be done.
It will never be perfect. There are too many similar faces.
I read once long ago, before the Internet, in some scientific journal somewhere that there are only a limited number of BASIC ‘faces’ for each race. I forget what the exact number was, a few thousand or so. But all ‘faces’ are just variations of the ‘basic’ face, some being more symmetrical some being less. Some being wider or more narrow, etc.
Now, if you are looking for an individual, and you have his or her EXACT facial measurements, eye width, chin length, ear separation, etc, it might become more accurate, but the computing power required would have to be enormous for it to be useful in real time........