'100% reliable' is brazen spin. The '100%' is, almost by definition spin and 'reliable' seems purposefully vague, as 'accuracy' is clearly the issue.
It doesn't appear that the technology is new here, just the expansion of the application.
The reality is, across 18 million people, most are going to have a lot of other people that look surprisingly close to them.
Face recognition would make a fine tool for (say) sifting through thousands of mug shots to match up an image of an unknown perpetrator to some candidates in what is really a large list of suspects (people with priors).
Right. That's what they are doing. They are not talking 'gen pop'.
Automated face recognition doesn't have the benefit of those other factors to help it narrow down it's search--rather it matches on anything that looks like the candidate image...
Just the opposite actually; since these are mug shots, they contain copious amounts data that can be used to limit the dataset for the actual facial recognition fuzzy image search. For instance, with your example of a local serial rapist the first order data set might only contain mugs of people who:
- Live within a 50 mile radius
- Weren't in custody at the time of the crime
- Have been arrested for rape in the past
- Meet the general Height, Weight, Skin Color, Age, Sex
What might that bring 18 million down to? 5000? 10000?
Or maybe search all mugs shots where ALPR had recorded a matching mug's plate at the approx. time and the place of the crime.
I think the main limitation is the quality of the photo from the crime, which obviously can vary greatly. Many pics are probably not worth the trouble today, but with the number of cameras increasing as well as their image quality, its value is growing everyday.