NIST study finds facial recognition algorithms struggle to identify masked faces

Examples
Date

A preliminary study finds that facial recognition algorithms struggle to identify people wearing masks. The study tested 89 commercial facial recognition algorithms, and the best had error rates between 5% and 50% in matching unmasked photos with photos of the same person wearing a digitally-applied mask. Masks both lowered the algorithms’ accuracy rates and raised the number of failures to process. The more of the nose is covered by the mask the lower the algorithm’s accuracy; however, error rates were generally lower with round masks; and the algorithms generally performed worse with black masks than with surgical blue ones. False positive remained stable or declined a small amount.
 

https://www.nist.gov/news-events/news/2020/07/nist-launches-investigation-face-masks-effect-face-recognition-software

Writer: NIST
Publication: NIST