That is, for now, 100% effective. I'm a former lead software scientist for one of the leading FR companies in the world. Pretty much all FR systems trying to operate at real time use a tiered approach to facial recognition. First detect for generic faces in an image, which collects various things that are not human faces but do collect every human face in an image. That's tier 1 image / video frame analysis, and the list of potential faces is passed on for further processing. This tier 1 analysis is the weakest part, if you can make your face fail the generic face test, it is as if you are invisible to the FR system. The easiest way to fail that generic face test is to not show your face, or to show a face that is "not human" such as has too many eyes, two noses, or a mouth above your eyes in place of any eyebrows. Sure, you'll stand out like a freak to other humans, but to the FR system you'll be invisible.

Wrt. cameras with depth sensors like face unlock this isn't supper likely to work.

Wrt. public cameras which don't have such features and are much further away and aren't supper high resolution either it maybe could even somewhat work.

There’s even a make-up trend of “enlarging” the eyes by painting the waterline of the lower eyelid white, that could be used as a justification for walking around like this even in a totalitarian police state.

Juggalo makeup is supposedly extremely effective.

Just make sure you don't how how magnets work for plausible deniability.