These are different problems, with different sensitivity/selectivity (and fail modes)

It's one thing to identify a person from a database, another from camera 1, then camera 2 etc. My guess is the latter has a much lower false positive rate