Look at a kid under the age of five, and it’s hard to imagine what he’ll look like in 70 years. But this new piece of software does just that. Check out this series of photos, which compares actual photos of a boy as he grows up (photos on the right) with photos generated by the new aging software, using only the three-year-old picture as a reference (photos on the left):
Of course, many computer scientists have tried to make face-aging software before. The umbrella field of getting computers to recognize human faces is a hot topic of research; Facebook recently published some work on getting its “DeepFace” software to recognize people from the side, given only head-on pictures. This new work is based on the largest-yet database of photos for aging software—40,000 pictures of people ages 0 through 100. The new software is also unusual in its ability to create accurate results from photos of very young children.
Software like this would be especially helpful to missing children searches, Seattle TV station KOMO News reports. Right now, expert artists try to help with searches by making drawings of missing kids at their current age. The artists use a combination of photos of the kids, the looks of the kids’ older family members, and current knowledge about how faces age. (Scientists already know, for example, that people’s faces and noses lengthen as they get older.) The craft of interpolating how a person will have aged is “part art, part science and a little intuition,” as one firm describes it.
This new software boosts the science part of that a bit. It’s based on measurements of about 1,500 people for every age group, including very narrow age groups for kids, who can change drastically from year to year. The photos came from, well, the internet. In a paper, the software’s creators—three researchers from the University of Washington and Google—described how they searched for photos to analyze:
Aha, so that’s who’s been looking at those old photos of you competing in Math Olympiad.
The team wrote algorithms that calculate, based on its database of photos, what’s different between photos of people at age X versus photos of people at age Y. How do face shapes develop between those two ages? How do skin textures change? The algorithms also deal with things like funny facial expressions and weird lighting that might show up in a reference photo. All the software needs is one photo of a child to create a series of images for ages up to 80.
One of the software’s creators, Ira Kemelmacher-Shlizerman, told KOMO News she contacted the Center for Missing and Exploited Children about her work. But the software isn’t ready for crime-fighting yet. The team wants to try to add other things to make it more accurate, including hair color changes and ethnicity-specific data, if that’s relevant. (To my untrained eye, the algorithm already appears to work well for people of a few different ethnicities.)
You can see many more age series like the one above in Kemelmacher-Shlizerman and her colleagues’ website and paper. They will present the paper at an Institute of Electrical and Electronics Engineers conference in June.