Researchers develop method to identify people even when faces are hidden

August 10, 2016 - 

German researchers have developed a new method that identifies people even when the majority of their photos are untagged or obscured, according to a report by Motherboard.

In a new paper, researchers at the Max Planck Institute in Saarbrücken demonstrate a method that takes into account previously-observed patterns related to a person’s head and body, to accurately identify the individual even when his or her face is blurred or completely hidden

This “Faceless Recognition System” works by training a neural network on a collection of images displaying both obscured and visible faces, then applies that data to intelligently predict the identity of obscured faces by pinpointing similar features found around an individual’s head and body.

The system’s level of accuracy depends on the number of visible faces that appear in the set of images.

For example, when there are 1.25 instances of the person’s fully-visible face the system has an accuracy rate of 69.6%. In cases where there are 10 instances of an person’s visible face, the accuracy jumps upwards to 91.5%.

Therefore, even if your face was obscured in the majority of your Instagram photos, the Faceless Recognition System would likely be able to identify you so long as there were one or two photos displaying your face in full view.

It turns out this becomes a lot harder to do using sets of photos from “across events,” or when factors like illumination and the person’s clothing have changed.

In addition, the researchers discovered that the system’s accuracy declines significantly from 47.4% to 14.7% when attempting to identify faces obscured by black squares across events.

However, this accuracy rate is still three times higher than the “naive” method of identifying obscured faces by means of prediction, researchers said.

In June 2015, Facebook head of artificial intelligence Yann LeCun revealed the company has been developing an experimental facial recognition technology that can use various visual clues such as hair color, clothing, pose, and posture, in photos to identify individuals, with 86% accuracy.

But the researchers assert that their Faceless Recognition System is the first to do so using a trainable technology that takes into account a full range of body cues involving blurred and blacked-out faces.

“From a privacy perspective, the results presented here should raise concern,” the researchers write. “It is very probable that undisclosed systems similar to the ones described here already operate online. We believe it is the responsibility of the computer vision community to quantify, and disseminate the privacy implications of the images users share online.”

Leave a Comment

comments

About Justin Lee

Justin Lee has been a contributor with Biometric Update since 2014. Previously, he was a staff writer for web hosting magazine and website, theWHIR. For more than a decade, Justin has written for various publications on issues relating to technology, arts and culture, and entertainment. Follow him on Twitter @BiometricJustin.