There have been pretty huge leaps over partially obstructed facial recognition algorithms in the last few years -- I think state of the art is approaching no meaningful loss of match mask vs no mask.
iPhone face unlock has worked while wearing a mask for a while now.
You know how you can identify exactly which family member is walking in the door from across the house, and maybe even what mood they're in? Or how little clarity you need to identify your child or spouse on a 480p camera feed? That's what machine learning makes possible across all types of sensor input, picking out those little but distinct patterns. There's really no way to be anonymous in public once ML surveillance software is widespread.
It is most certainly not, unless you’re talking straight on well lit training data.
https://www.mdpi.com/2076-3417/11/16/7310 https://www.sciencedirect.com/science/article/pii/S240589632... ...
That said I have no insight as to how many of these techniques have been found to scale well or have started to make it into product. It has been publicly reported that NEC’s NeoFace (a system that many police and govt use) newer versions does indeed have occlusion (mask) recognition operating at very high levels.
anyways thats just my understanding as an interested bystander -- not in the field.