Standing on the RIM of greatness
Your brand is your user experience
Are you obsessed with Pinterest?
Big Orange Valentines
If you haven’t tried the facial recognition tools within Facebook, look for the privacy setting with the following label:
“Suggest photos of me to friends.
When photos look like me, suggest my name.”
Because I’m a programmer, I felt compelled to do a little more digging on facial recognition (and the opportunities that may come from it). As it turns out, Face.com, (the company supporting this feature) has an open API for facial recognition, so I decided to run some tests using the first 300 pictures from the Grip Xmas party as my data set. For those interested; I’m using a very low tolerance level (20-50%) in the searches I’m running. If needed, you could dial up the tolerance levels for more accurate results.
Experiment 1: Image Grouping – finds groups of reoccurring faces in the data set.
Pretty impressive for a first scan. There were some errors, but it’s interesting to note that the system can also learn from its mistakes, and most of the incorrect matches came out better after invalidating the matches and running it a second time.
Experiment 2: Face Analytics – analysis on faces found in the data set.
It correctly picked up that:
a) Brian and Jacoub are both guys.
b) Brian doesn’t have glasses but Jacoub does.
c) Brian isn’t smiling, but Jacoub is.
I’m also tempted to think that its algorithm doesn’t take facial hair into the equation, as some of the accuracy rates are way off in several pictures for anyone with facial hair. So perhaps it’s best to avoid using this feature during Movember, or during an international beard growing championship.
Did you notice the extra person in picture by the way? I didn’t, but the computer did. I thought it was an error at first, but when I cranked the levels up it showed me this:
Pretty cool, huh?
Experiment 3: Crowds.
You aren’t able to submit images over 900 px wide without a premium service, but despite a fairly low image resolution, the scan still managed to find a lot of faces. Unfortunately it was only able to provide metrics for a few of those faces such as Ashlea (the person on the extreme left of the image), who happened to be looking in the direction of the camera.
Errors
Granted, it’s not without errors. After all, the metrics on Jacob (in the foreground) show only a 20% chance that he’s a girl.
Given that this landscape is pretty open, and the extreme popularity of photo collecting and browsing, this software could open up some pretty fun creative channels.
What do you think?
Jon Finkelstein
January 11, 2011 @ 12:31 pm
Great article, Raz.
I KNEW there was 20% chance Jacob was female!
Jacoub Bondre
January 11, 2011 @ 3:19 pm
Also not there is a 0% chance he is male, and an 80% chance he is not a human.
Its amazing how accurate these computers can be these days.
Tweets that mention Diving face first into creative opportunities « Big Orange Slide -- Topsy.com
January 11, 2011 @ 3:44 pm
[...] This post was mentioned on Twitter by Jacoub Bondre, Victory. Victory said: Facebook has it? Creepy! RT @running_cloud How accurate is facial recognition? http://bit.ly/ghv4gl [...]
Jacob
January 11, 2011 @ 3:55 pm
Interesting stuff, there is a nice free software out there that does this kinda of image, not just facial, grouping. i have used it on a few websites. very handy.
(would post link but not sure if this blog allows that).
Gil Hirsch
January 12, 2011 @ 2:36 am
Great read Raz! I’m Gil, CEO @face.com, and this feedback is super helpful to us, especially as we’re phasing out of alpha.
Drop us a line at our forums or gil@face.com and we’d love to help you make the most of our APIs. We’d also set you up with higher rate limit accounts.
gil@face.com
JamesBang Ryu
February 7, 2011 @ 2:50 am
Wow~ It looks great.
Melanie
February 8, 2011 @ 3:17 pm
cool – maybe we should submit the Mona Lisa painting and know at last if she is smiling or not!