If you're a parent wondering what your child will look like as an adult, now you don't need to wonder anymore. Researchers at the University of Washington claim to have developed software that can accurately predict what a child will look like as an adult, up to the age of 80. The technique can even work from poorly lit photos, and could prove a big help in missing persons cases.
Predicting what a child will look like as an adult from a single baby picture is extremely difficult, even for the people whose profession it is to create facial composites for the police. Try to get a completely automated software to do the same, and you'll run into an additional layer of problems in the source picture, including poor lighting, strange facial expressions, and other imponderables like, say, a milk mustache.
But researchers led by assistant professor Ira Kemelmacher-Shlizerman say that they've managed to create a computer algorithm that can actually do a much better job than its human counterpart.
The software relies on a database of thousands of faces, grouped by gender and age. For each age bracket, it determines the average pixel distribution of various facial features, and then calculates how those features change from one bracket to the next, as age advances.
The system's algorithm then takes a baby picture as input. After correcting for tilted faces, turned heads and inconsistent lighting, the software applies the age bracket-related changes to the child’s photo to predict how he or she will look like as an adult, reportedly guessing his or her appearance with remarkable accuracy, up to the age of 80.
In order to test the algorithm's performance, Prof. Kemelmacher-Shlizerman and colleagues collected the actual pictures of 82 people over a span of several years, and fed the subject's baby pictures to the software. From that single input, the program-generated facial composites were so close a match to actual photos that when a group of people was presented with the original and the rendered pictures side by side, they were unable to tell which was which.
The system could prove extremely useful in missing person cases. Police facial composites are usually created manually, taking photos of the child and family members and manipulating a best guess of how the face may have aged over the years. This algorithm, on the other hand, has proven remarkably accurate and, run on a normal computer, it provides an answer in about 30 seconds.
A freely available paper (PDF) published by the researchers explains in detail how the algorithm works. Prof. Kemelmacher-Shlizerman told us that the technology will soon be made available to the public.
The video below shows the software in action. On the left is the original photo, while on the right the algorithm outputs the prediction of what the child will look like as an adult.
Source: University of Washington