If ML doesn/t work well... you need more data!
I can see the quandary here, which is that these facial recognition systems will continue to be inaccurate without more training data.
But the easiest way to get more training data is of course to deploy the system in public.
So those arguing that inaccuracy is a reason not to deploy the tech in public are missing the point, or perhaps being disingenuous; the tech will never become more accurate without public deployment.
The trouble with focusing the debate on accuracy is that it won't be a problem forever, you're only delaying the inevitable deployment of a system trained with data gathered elsewhere or by other means.
If you are opposed to seeing any kind of live facial recognition on the streets, accurate or not, you need to have the courage to argue against it from a morality/privacy standpoint rather than focus on accuracy, ethnic bias or any other problems that are inherently solvable.