Researchers use Instagram mask selfies to improve biometric facial recognition algorithms

Some facial biometrics developers fear the COVID-19 pandemic will affect their business, as the number of people wearing masks has increased. Because they cover essential facial features, masks make it harder for algorithms to recognize users, so researchers need data to improve them.

Researchers are collecting social media photos of people wearing masks, without their consent, to feed them into biometric facial recognition algorithms to improve detection and accuracy, writes Cnet after identifying thousands of photos, mostly taken from Instagram, available in public data sets. The COVID19 Mask Image Dataset  published on Github in April had over 1,200 pictures taken from Instagram. It used AI startup Workaround to filter the images.

“We were inspired by all the companies that were launching free tools and everything they can do to help,” Workaround CEO Wafaa Arbash told Cnet. “We have these public images from Instagram, so these are not private images. We were just searching and getting the right data.”

A number of facial recognition companies have asked employees to send selfies wearing masks or digitally add masks on existing photos, which is how the U.S. National Institute of Standards and Technology (NIST) plans to test the technology.

NIST has announced a series of tests for face mask effect on facial recognition accuracy. The first step will be to digitally add synthetic masks to faces and test 1:1 verification algorithms already submitted. New algorithms can also be suggested for testing.

“We will first mask only the probe image, leaving the reference photo as is.  Later, we will consider the effect of masking both images. We will quantify the effect of masks on both false negative and false positives match rates and issue a public report,” the organization wrote.

For their research, Arbash explained the Instagram photos were found using mask-related hashtags. They had initially collected 3,000 pictures, but reduced the list to 1,200. Cnet writes one of the pictures included a child, yet Arbash said it may have been an error. The people in the photos were never asked for consent, as their profiles were public and not set to private, Arbash said.

“We’re not making any money off of this, it’s not commercial,” Arbash said. “The goal and the intention was to help any data science or machine learning engineers who are working to fix this issue and help with public safety.”

Researchers at Wuhan University in China have allegedly created the Real World Masked Face Dataset which contains over 5,000 photos of masked faces “from massive internet resources.”

Face mask detection technology is a priority for developers, yet privacy advocates have voiced concerns regarding the methods used to compile the databases.

Article Topics

accuracy  |  AI  |  algorithms  |  artificial intelligence  |  biometric testing  |  biometrics  |  biometrics research  |  dataset  |  face photo  |  facial recognition  |  mask detection  |  NIST  |  privacy  |  training