Google Blogoscoped spotted
the Google Image Labeler
game, designed to help Google improve its image search results through tagging.
It feels like a catch-up game with human-powered efforts that Yahoo is embracing
via Flickr — plus it also looks pretty influenced by the work of Luis von Ahn and his
The game pits you against
someone else. If you see a picture of a car, and you both label it car, you can
proceed to the next image. You continue until your time has run out. Here is an image of Horcrux and
Barry Schwartz from our blog (rustybrick) scoring 300 points for matching tags
on three images.
Image search has been tough
for search engines. They can’t easily tell what an image is about, since there’s
no good way to "see" the images and categorize them. Some technologies to
recognize faces, colors, shapes, objects and other things are improving. Still,
it’s hard with an image of someone like Martha Stewart. Is she a woman,
celebrity, criminal or just Martha Stewart? Or all of these? How do you know
which one or ensure that all of them are applied.
Yahoo’s Bradley Horowitz is probably one of the most famous converts from
turning to human power over computer power. He’s been cited many times as having
originally sought a technological solution to understanding what’s in video and
image data, then moved to embrace people power. Here’s one example of that from
a Wired article
Horowitz’s favorite project is incorporating people-powered metadata
systems from two other Yahoo! properties: the recommendation technology from
Yahoo! Music and the tagging features from Flickr, the photoblogging company
Yahoo! acquired this spring. Google’s original stroke of genius was figuring
out how to piggyback on human judgment by following the links people make
between Web sites. Horowitz is borrowing functionality from two Yahoo!
properties to develop something similar for video.
We’ve just seen Yahoo make more of a commitment to using that human power
when it started inserting Flickr results,
rather than Yahoo Image results (which are computer sorted), into regular web
searches last week.
Google, of course, has no Flickr to use. Enter the game. It’s designed to get
lots of people to quickly label images because they want to have fun. If that concept
sounds familiar, it’s because that’s exactly the method behind the
ESP Game, created by Carnegie Mellon
professor Luis von Ahn.
I first learned of Ahn’s work through a 2003 Associated Press article,
Researchers Hope to Improve Web Searches. His ESP Game came later. The
Google system looks like a copy of it, perhaps with his cooperation. Perhaps
he’s even there now — I’m checking. We know he was there just last month,
here’s a video of his lecture on classifying images (well worth watching). That Martha
Stewart example above? That came from his video. I’ve also embedded it below:
In his video, Ahn thinks that in two months, all images on Google Images could be
labeled. It’s not hard to believe after hearing that, Google jumped to try this.
Postscript: I heard back from Luis von Ahn, who sent me this:
Yes, Image Labeler is based on my ESP Game, which Google licensed. I’m not
employed by Google, however, since I’m a full-time faculty member at Carnegie