Wednesday, December 14, 2016

Averaged faces aid recognition

Below are some close-up images of John Travolta. Which one is the most "classic" or identifiable as Travolta?

Every photo presents unique variations of hair, expression, angle, age, and lighting. Sometimes the person hardly looks like themself. If you have ever painted a portrait likeness, you know that it helps to have a lot of photos of the individual. If you copy just one piece of reference, you may not get a recognizable likeness at all.

So which one is the classic Travolta for you? If you chose the one in the center, there's a good reason for it. It's not a photo at all, but rather a computer average of the other photos. 

Computers can take a set of photos of a person and blend them into a single average face, erasing vagaries of illumination, perspective, and expression. 

Here is a set of those averaged faces of celebrities. They have polygonal borders, but those polygons at least keep some sense of shape and proportion. To my eye, they're almost all instantly recognizable (unless I didn't know them at all). I'd guess that they're probably more identifiable than any single photo taken at random. 

The scientists who did this work discovered that these averaged faces are also far more recognizable to A.I. facial recognition systems than are random photos. As the authors put it "the simple process of image averaging can dramatically boost automatic face recognition." Sometimes the averaged faces boosted the success rate from 40% to 100%.

Read more at the free PDF:


krystal said...

Cool! Sat in a facial recognition workshop earlier this year, using the Azure Face API from Microsoft. You can probably try it out just by using a trial account, (it's .Net, though, so you'd probably have to know a little bit of that platform, but it's nothing that can't be learned if you're dedicated!) etc. It was great in also accounting for things like depth, so it's not just recognition from a distance-in-2D point of view, but also from a depth (like checking for things like perspective, to tell the difference between a photo and an actual person) point of view. Thought that was neat. The (Microsoft technical) evangelist who was doing a demo for us at the time was doing a system (it sounds like a terrible idea in retrospect!) where you'd have to verify picking up your child from daycare with both voice recognition and image recognition, using the voice api and the facial recognition api. It didn't work :( but.....(disaster also if say, someone else has to pick up your child in case of an emergency, and well, multiple other scenarios).
I'm also really interested in OCR stuff!

Bill Bridges said...

Very cool! Thanks for posting this. To me, you can even eliminate the mouth from the image and they are recognisable.