Hi all, I'm a research engineering manager on the Microsoft Project Oxford Face API team, thanks for the interest!
This website makes use of the Project Oxford Face API technology we announced yesterday (http://www.projectoxford.ai). The age prediction is an experimental feature based on MSR face research investment. With Project Oxford's beta REST API we make this available to developers. These models will improve over time, please keep the feedback coming.
Did you try it? Because in using several pictures of me and other people I know it seemed wholly inaccurate. If they were cheating, they're not very good cheaters.
I thought it was doing this with the first picture of myself I tried, because it was a picture from a few years ago that guessed my current age. After trying another, 5 year old picture that is the portrait I use for social networks and it guessed that I was younger I figured this wasn't the case :).
I used two photos of me both taken the same day (corporate headshots that both appear on wedsites and in social media with my name), and it returned different answers. Both answers were significantly off.
I am often told that I look very young for my age, and people generally guess my age several years lower than it is. This tool guessed the correct age right away, which really impressed me. Thanks for the confidence boost Microsoft!
The accompanying article is also very interesting and well written to be readable by someone who isn't totally up to date with the latest in facial recognition software.
I uploaded a photo with a picture of myself (I'm 25) and my 11-year old sister. I got guessed as 44, she was guessed at 74.
I know I look older than my age (though I wouldn't think 19 years older), but my 11-year-old sister coming up as 74 was surprising. What's 63 years, give or take?
"I am often told that I look very young for my age, and people generally guess my age several years lower than it is."
Ditto, except this tool suggested I was three years older than I was at the time. Of note is that the photo was taken three years ago, so it's correct now. However, that was pre-beard, which generally ages people, so ... I don't know what to think.
Same here; though the guess was nine years older than I am! I am just going to assume that they don't have enough folks that look like me in their dataset.
I'm curious to hear more about the accuracy of their machine learning algorithm though. I'll need to read into that later.
Same here, but I'm 25 and the tool guessed 20,21,27,27. I guess the average is correct-ish.
In my case, the old guesses are when there is bright light or when my head is slightly turned. The younger guesses are with regular room brightness 100% front face.
I think it might not be entirely clear what this site represents. It's actually quick demo of ML services available on Azure. I think it uses Deep Learning and the power of it is to demonstrate that you can wire up something like this yourself without write a line of code for deep learning algorithms or even owning any servers at all for heavy GPU processing. I wish they had put code for this website on Github so people can tweak and spawn new versions.
The results on every photo I tried were ludicrously wrong. The closest it got on photos of me was about five years older than me, but it ran anywhere up to 20 years too old.
The same thing on photos of others I tried: a 22-year old man was estimated to be 49, a fifty-year-old woman--who is generally considered to look young for her age--to be 69. Those are ages that no one would ever guess based on looks.
So I'd say the algorithm needs a little more work.
There are dozens, maybe hundreds, of pictures of me on the internet already. I certainly don't care if this site has one, and I couldn't care less about its transfer being encrypted.
This. My girlfriend was happily uploading picture after picture until I asked her semi-rhetorically "I wonder what they're doing with all those uploaded pictures?" and after thinking about that for a second she closed the tab.
>In addition to age and gender, we also got additional information such the User Agent string of the users of the web site, the latitude and longitude of location from where the picture was uploaded and more. This is represented in following JSON document.
Seems like they really don't need to know where you were when you took the picture. But they gather that data anyway. :)
While I absolutely agree with the sentiment, I must point out that they accidently have SSL, even if they don't know it, because CloudFlare. It's just not forced, which is unfortunate. And yeah, the lack of privacy stuff is concerning.
I have a big ol' beard and it guessed my age (32) to within one year. It thought my wife and sister were 10-15 years older than they really are, and it thinks all my kids are girls.
Looks like this is a demo for the Microsoft Azure ML. The main value-add here seems to be integration with the rest of Microsoft's technology suite, with its visualization, live streaming stuff, etc. For the actual age/gender estimation, I'd be curious to see if it's much better (or better ta all) compared to OpenBR (http://openbiometrics.org/).
They should really run this against a bunch of photos from a dating site to measure the difference between the claimed age and guessed age and then compare to the result from whatever training set they used. I'm guessing they would get a bimodal distribution for dating site photos -- some people claiming to be younger than they are, and some people giving their actual age but using very old photos.
I'm at a family thing so just tried it on about 5 people in a row. My two daughters it got 100%. My mother in law was guessed to be 10 years younger than she is (she was happy). My father in law was spot on (early 60s). This is pretty impressive. (I should note that it was VERY bad with dark/poorly lit shots with the eyes hard to see so we moved to a naturally lit space.)
From one photo it guessed 24, from another it guessed 52, both less than my true age. I'm sure it will improve, but that's not very impressive.
Having said that, assuming it's fully automated, it is actually impressive that it should be attempting anything like this at all. Not that many years ago it would have been completely infeasible to anything of the sort.
Hello, I am a program manager at Microsoft. I would like to point you to the Azure Machine Learning Gallery (
https://gallery.azureml.net/).
There you can find both Machine Learning API (including the one used in how-old.net). We have many services that allow you to harness the power of machine learning: Speech APIs, Recommendation Engine, Text Analytics, Customer Churn Prediction APIs, etc.
I'm in my early twenties, as are most of my friends. But it consistently guesses we're 10-20 years older than that. I wonder whether that's to do with the population it was trained on?
OTOH, we're all PhD students, so maybe we're all just stressed out and look old.
Interestingly, I have long hair, which I think is why it sometimes misclassifies me as female. And when it does that, it seems to consistently give me a younger age.
That's ok, I'm 37 and it guessed I was 81 in one photo and in two others, it couldn't detect my face at all. To be fair, I was wearing a bike helmet in the 81 image, sunglasses in one failed image, and a baseball cap in the third.
OTOH, my wife will be pleased to know it underestimated her age by a decade.
If you're interested in the human side of age perception, we're running a study at http://testmybrain.org ("Understanding other people") where you judge people's age based on a photograph. We're looking at individual differences in face perception — how your age, race, and experience affect your judgment of others.
The quality of the prediction also matters on the quality of the photo. I tried the photos that I do have of myself online, which each have a distinct lighting profile. One is a soft, orange glow in a restaurant... the guess was off by +11 years. The next photo has a portion of the left side of my face obscured by shadow. This was off by something like +31 years.
I don't get it. When I click on "Use This Photo", it always says, "Couldn’t detect any faces. Please verify that the image is valid and less than 3MB." When I type anything in the search box, it always says, "Oops, something went wrong. Please try searching again."
2 most recent, one with beard and one without are both 1 year older than I actually am... not bad.
One from 4 years ago was guessed at one year younger than I actually was, again, not bad. But according to this, I've aged 6 years in the last 4. See what programming will do for you?
Just going by the test images it seems to rate everything about 5 years younger than what I would guess (people and their kids, kids being the same age and obviously not twins).
With that said, it guessed my age nearly spot on. When I get carded most people think I am far younger.
Everyone here is getting judged younger; I must be the only one who it thinks is actually older. I am 34, but it thinks I am 36 to (in very bright lighting) 56. It seems to be closest to believing that I'm 39.
I took three pics of myself with Photo Booth just now, and depending on the lighting and position of my chin got 20, 34, and 47. I'll take the average?
FWIW, I'm a 22-year-old man, and it consistently puts me in my 40s. I might look a little older than I am, but not by that much... I suspect you and I (and many others) just have faces that don't agree with the classifier.
Ray Kurzweil current photo - 39 (actual age 67). Damn, those meds are working (no sarcasm here, I think the overall accuracy of this tool is stunning).
It got the picture of me and my cats almost spot-on: 32. Since I was older than 31 1/2 at the time, I'm going to call that a hit. I attribute that to the cats.
I think I win for maximum error, though. On a picture of Jeanne Calment on her 122nd birthday, it guessed 43. Off by 79 years.
