In the ideal world, forensic face specialists benefit from having all available facial data to perform their examinations.
But what if they don’t have a full view of the faces?
And do they even need it?
When masks were the rage
The biometric world encountered this issue a few years ago, when a worldwide COVID epidemic caused people to wear the newest cool accessory. Face masks that, when worn properly, covered the nose and the nouth.
The facial recognition algorithms at the time assumed the continued visibility of unchangeable facial features. The eyes. The nose. The mouth. Take some of those away, and accuracy decreased.
So NIST sprung into action and began a new test. Beginning in March 2020, NIST took existing facial images, superimposed face masks on them, and tested the accuracy (1:1) with the face masks. As expected, newer algorithms developed after COVID were more accurate with masked faces. NIST discontinued this particular test in February 2024.
When parts of a face are covered
After NIST’s face mask testing ended, researchers performed a different test in which “participants matched an occluded probe image to one of two target identities by selectively revealing isolated facial features (e.g., ears, eyes, nose).” Kinda like slowly revealing the pieces of a puzzle.
“We focused on nine facial features: cheeks, chin, ears, eyes, eyebrows, forehead, neck, nose, and mouth. These align with parts of the face described in FISWG [Facial Identification Scientific Working Group] guidelines and prior research.”
To conduct the test, the researchers obscured the probe image completely, and test participants “were allowed to reveal one feature at a time on the probe image until they were ready to make their decision.” Kinda like “I can identify this person with 2 of 9 features visible.” The goal was to keep the number of features revealed as low as possible.
After the participants had finished, the researchers examined the resuilts. They found: “Trials where only a single feature was revealed were the most frequent occurrence, comprising 396 out of a total of 648 trials.”
Read that again.
Of all nine features (cheeks, chin, ears, eyes, eyebrows, forehead, neck, nose, and mouth), over half of the tests completed after the participants had revealed only a single feature.
And which single features were used for many of these identifications? The forehead and the ears.
“Foreheads were associated with the highest accuracy (probability of success = 0.94), though the lower bound of the 95% confidence interval reached a comparatively low value of 0.71. In contrast, ears showed a similarly high probability of success (0.93), but with a tighter confidence interval (lower bound: 0.88).”
Now let me caution that these were people and not algorithms, so this is not really comparable to NIST’s face mask testing. But it suggests some ways to improve algorithms. If they’re not already doing so, perhaps the algorithms should more thoroughly analyze foreheads and ears. The eyes are apparently less important when real people do the work. And all available facial data may not do the job.
