frvt

If the best face recognition software has an error rate of 0.08% [1] and Rite Aid serves 1.6M customers a day [2] what moron thought that was a good idea? [1] https://github.com/usnistgov/frvt/blob/nist-pages/reports/1N... [2] https://careers.riteaid.com/pages/company-info

This is wrong and not how the numbers are computed. It's not pairwise. Pairwise error rates are more like 1 in 10 billion. NIST did a review in 2019 (published 2020) with datasets containing 12 million people. For searching a given photo against the database, solutions might, say, fail to pull anyone 3% of the time and pull the wrong person 0.5% of the time. Thus, you get about 1 false positive for every ~190 true matches. It was also under pessimistic conditions: a big fraction of the searches were for individual not enrolled in the dataset. https://github.com/usnistgov/frvt/blob/nist-pages/reports/1N... edit: I see there are newer versions of the data since I bookmarked this, but I have not reviewed them. A glance shows they format the data differently. Still, things have likely improved somewhat since then.