The system I worked on had (it's larger now) 100M faces in the training set, and when I was leaving there was a 300M set in the works. We went to lengths to collect and categorize. It's mixed ethnicity that throws a wrench into ethnic categorization. It is ordinary to have people with a half dozen or more racial compositions, and that pretty much wreaks categorization. We (they) also had a a pretty robust liveness detection, and surgical mask detection with a "see through the mask" feature too (available with certain crazy-tech cameras.)