Microsoft Face Rec App Guesses Your Age

Microsoft has released the free 'How Old?' webpage that guesses both the gender and age of the picture's subject.

I tried it out and got mixed but fairly accurate results, ie:

Indeed, I am both male (indicated by the symbol next to the age) and 36 years old. I tried it with a mix of images, and it was fairly consistent. It never got my gender wrong (thanks goatee!) and it generally guessed me at 37 years old, which is pretty close.

However, not everyone experiences good accuracy. From comments on Microsoft's Face API page:

The girl above is obviously not 56. Nor a man.

And this girl is clearly not a 74 year old man, either.

Many users report similar errors, so it is not foolproof by a long measure. Images with groups of people don't appear more prone to error, but some faces are consistently tough to 'guess'.

However, the analysis works on faces previously unregistered with (unknown to) the database, so it is interesting.

Demographics analysis has been a niche feature of analytics providers for a while, ie: 3VR Demographic Analytics (Gender / Age), so Microsoft's implementation is sure to attract interest.

Give it a try and let us know how accurate it is for you.


Sorry Bruce Jenner. Microsoft says you still have a ways to go.

It thought I was 36 :( But I'm only 35!

Does this app go below the 30's? I'm 25, but still, not a single correct age guess (and it thought I was a female in one image)...

That's an interesting and hilarious result, Derek.

Suggest IPVM automatically rank shootouts based on how accurately each camera 'guesses' Derek's age.

Manufacturers may subvert this over time by modifying their firmware to recognize Derek's facemap natively, but this will likely take a while.

Few comments:

From what MS claims in their blog, they were really just looking for feedback and only initially released this to 50 people before it went viral. (Fun with ML) Not sure how much I believe that given how polished the interface is and how well it scales under load the last couple of days..

There are some new developments in the ML technology that really do show promise for better machine learning in the near future. And the capability is more accessible than ever: go rent a rediculously powerful g2.8xlarge AWS instance for $2.60/hour.

In our business I believe the MS example shows how silly video analytics for *precise* answers can be. But I also believe that ease of access to to this increasingly inexpensive technology will enrich systems with features that have value albeit somewhat indirect--such as 3VR's model that bins faces into demographic categories could provide benefit to a retailer--just without the premium 3VR pricetag.