A new report published by the National Institute of Standards and Technology (NIST) shows the results of 65 new facial recognition algorithms created after the start of the COVID-19 pandemic, as well as 87 algorithms submitted pre-pandemic. The report revealed that software developers are getting better at developing algorithms that recognize masked faces, even getting as accurate as regular facial recognition algorithms. “While a few pre-pandemic algorithms still remain within the most accurate on masked photos, some developers have submitted algorithms after the pandemic showing significantly improved accuracy and are now among the most accurate in our test,” the report reads.
What the Study Found
The study was the second of its kind conducted by NIST with the same dataset meant to test facial recognition algorithms and their accuracy in the presence of face masks. The report’s authors used 6.2 million photos and applied simulations of various digital mask combinations to these images. Mei Ngan, a co-author of the report and computer scientist at NIST, told Lifewire in a phone interview that the presence of face masks has essentially taken facial recognition technology back about two to three years. “Error rates are anywhere between 2.5% and 5%—comparable to where the state-of-the-art technology was in 2017,” she said. A previous report from NIST published in July looked at the performance of facial recognition algorithms submitted before March 2020, prior to the World Health Organization declaring a global pandemic. This first study found the error rate of these pre-pandemic algorithms to be between 5% and 50%. Even if these algorithms are getting better at reading masked faces, the more recent study found that some factors affect the error rate, such as mask color (darker masks like red or black have higher error rates) and how the mask is shaped (rounder mask shapes have lower error rates). Ngan said the algorithms use the visible part of someone’s face, such as the region around the eyes and the forehead, to recognize facial features rather than read through the mask itself.
The Future of Facial Recognition and Face Masks
Ngan said it’s obvious developers have made significant improvements with their facial recognition algorithms when it comes to face masks. “There is clearly a need for facial recognition systems to operate under the constraints of wearing face masks,” she said. “Given the stuff we’ve been doing and the outcomes from our recent study, we are seeing the facial recognition industry is actively working to include face masks in their algorithms.” Since the technology is improving, that means it’ll be easier to do things like unlock our phones while wearing a face mask, but there are other implications when it comes to facial recognition advancing in this way. Numerous studies show that facial recognition is widely reported to misidentify the wrong person and have racial biases. A 2019 study by the NIST found that facial recognition technology misidentifies Black and Asian people up to 100 times more often than white people. Even if the technology is getting better at reading face masks, the error percentage—no matter how small—could still be a concern for misidentifying a person wearing a face mask. While the most recent NIST report shows that algorithms are getting better at handling the face mask task, Ngan said only time will tell if this really is where the future of facial recognition is going during pandemic times. “Maybe we can expect further error reductions, or maybe developers might find limitations to the amount of unique information in the unmasked region,” Ngan said.
