Timœ Goes By: Verifying the Age Estimator How-Old.net

how-old.net is a project from Microsoft that guesses the age of people on a photo based on their faces. I have tested it with an image that consists of 21 photos. Each photo was taken around the time of my n-th birthday, so the ideal classification would label them as 0, 1, 2, …, 20. Let’s see what how-old.net comes up with.

But first, here is the image that I have uploaded. The first row goes from age 0 to 6, the second from 7 to 13, and the last from 14 to 20.

And these are how-old.net’s guesses:

The average age-estimation-error is about three and a half years. The gender detection in turn has failed for most of the photos. At the age of 0, well, I was not recognizable yet. It’s interesting to see that how-old.net was capable of making a relatively clear distinction between toddler (1 to 3), child (4 to 12), puberty (13 to 16), and adult (17 onwards).

There are three age estimations that are exactly correct. Counting the number of photos for which the estimation was off by at most one year gives seven out of 20!