How-Old.net service determines age from a photo
How-Old.net service determines age from a photo
Anonim

This fun project from Microsoft uses machine learning to calculate the age of users from facial features.

How-Old.net service determines age from a photo
How-Old.net service determines age from a photo

How-Old.net does not yet claim to be an accurate tool and exists rather as a free entertainment project. Visitors upload their photos and share the results to impress friends, or to amuse them when the service goes wrong. But How-Old.net is based on the power of the Azure cloud platform, and its algorithms are constantly being improved.

The site is convenient to use both on a computer and on a mobile device. The developers have added several template photos of smiling people, on whose faces you can test the accuracy of age determination. You can also experiment with celebrity images or of course use your own. Microsoft says it does not store uploaded photos.

Image
Image
Image
Image

As for accuracy, in my case, out of ten test shots, the algorithm was able to correctly determine the age of only four of them. But the error of the service turned out to be small: the robot added or lost only 2-3 years.

Try How-Old.net →

Microsoft went beyond the site and added support for the technology behind How-Old.net into its own Sprinkles mobile app. This program can detect emotions and apply decorative elements such as hats and masks to users' faces. Alas, while Sprinkles is only available for iOS devices connected to the US App Store.

Image
Image
Image
Image

It should be noted that this is not the first implementation of such an idea. For example, cameras of some Xiaomi smartphones have long been able to determine the age and gender of a person from a photograph.

You can also find a huge number of apps with similar features on Google Play and the App Store. But I could not find worthy ones among them. They either do not determine the age from the photo at all, despite the description, or they do it very badly.

Recommended: