Facial Recognition Designed For Face Masks Fails
Facial recognition software has come a long way. In fact, it's a bit scary, how good it has become - until COVID-19 and the face mask came along.
A new study shows that wearing a mask creates a lot of errors, even in the best software.
The US National Institute of Standards and Technology tested 41 facial recognition algorithms that had been updated after the COVID-19 pandemic. These companies hopped that the new method would not get stumped by a lower face mask. In fact many of these companies claimed that face masks could not and would not fool their new programs.
Well - sorry but - every algorithm experienced increased error rates when masks were introduced.
These programs still had rather good accuracy overall. So they don't totally suck. One company from China, named Dahua's, had an error rate of just around 0.3% without masks to 6% with masks. That's not bad. But other companies had a failure rate up up to 99%. Now that sucks.
Rank One ran an error rate of 0.6% without masks but 34.5% error with. That destroys their previous claim of being able to identify people just off their eyes and nose.
TrueFace is a company popular with schools and the Air Force. They went from 0.9% error without to 34.8% error with a mask.
These companies have gone back to the drawing board. Their new goal is to pick out a person with only a small portion of the face.
That would make this technology more unsettling than ever.