And Maw! This

Google Image Labeler

September 5th, 2006

I’ve heard a lot of talk from people like Danc, David Edery and Action about using gameplay to improve productivity and make work and learning fun, but seen very little results beyond pie in the sky ideas. Google Image Labeler makes good on some of those promises and whacky ideas. Turns out that this is really just Google’s implementation of the ESP game.
Humans are much better at labeling images than computers are, but we don’t like to work without some incentive. Lucky you can give people stupid rewards like points and high scores, and they’ll gladly do any repetitive task that isn’t too hard until they get bored with it. The game is simple, you have an unknown partner and you’re given 1:30 minutes to come to a consensus about how to label as many images as you possibly can in that time period. The system is easy to game as demonstrated by this thread on QT3. Another problem is that people go for the lowest common denominator because speed is rewarded over accuracy. A picture of a specific person usually becomes “man” or “woman”, pictures are often labeled “sky”, “space” or “trees” and if all else failed labeled “red”, “blue” or whatever the predominate color is. There are off-limits words to filter commonly used labels but it doesn’t work very well right now, often filtering obvious but sufficiently detailed words and not stopping people from using vague descriptions.

I would like to see more varietions on this theme. One could be a matching game where you are given an image and have to choose the most similar image from a list. If you and your partner agree on the image you score points. In later games matched images would be grouped together and people would choose from images that are more similar. Although it would start out random it (might) rapidly converge. It’s nice to see Google using games to get people to do their dirty work for them. It might end up useless unless you are searching for “man”, “woman”, “trees” or “sky”, but hopefully it will evolve into something that improves image search.

Entry Filed under: