Wednesday, June 27, 2007

Free human grunt work

It’s pretty hard to sell your fine art or stock images if no one can find them. In order for Web search engines to find your images, someone has to categorize and label (keyword) them. Although computers have taken over a lot of our repetitive tasks, there are some things computers haven’t learned how to do very well, and one of these is image recognition.

At last, someone has found out how to get humans to do boring “grunt” work without paying them anything—make a game out of it! That’s genius behind the Web hit known as the ESP Game, the brain child of Luis von Ahn, a 28-year-old professor at Carnegie Mellon University.

The Strategy. The ESP Game is a two-player Web game. Each time you play you are randomly paired with another player whose identity you don’t know. You can’t communicate with your partner, and the only thing you have in common with them is that you can both see the same picture. The goal is to guess what your partner is typing on each image. Once you both type the same word(s), you get a new image. Each time you type a word or phrase, you must press enter on your keyboard to submit it to the game. You can type as many guesses as you want, and as soon as a single guess of yours is equal to a guess that your partner has made, you get a new image. You have two and a half minutes to agree on 15 images. The strategy is to type as many words as you can related to the image, hoping that your partner types one of them. Each time you and your paired partner agree on a word or phrase, that word or phase becomes part of that image’s list of “taboo words” that the next group of paired players cannot use.

The Reward. You get points each time you agree on an image with your partner. The exact number of points depends on how many taboo words the image has: the more taboos, the more points the image is worth. If you agree on 5 images you get a bonus of 350 points, if you agree on 10 images you get a bonus of 850 points, and if you agree on the 15 images you get a bonus of 2000 points. You can check how many images you have agreed on by looking at the progress meter at the bottom of the screen.

People have been known to play for hours. Professor von Ahn knew that people who play computer solitaire will happily diddle away countless hours at their machines, and by adding a scoreboard to the ESP Game he knew he could tap into the competitive juices that drive people’s addiction to all forms of gaming.

Why? What is really going on here? What’s really happening here is free labor. Tom “ESP Game” Sawyer is getting you and I to white wash his fence, and making us think it is fun. ESP Game is taking previously unlabeled image collections and giving them keywords that can be used to categorize and retrieve them via a Web search engine! And because of Professor von Ahn’s clever programming, the buyers of his service can be assured that these are very relevant keywords, thoroughly tested by the competitiveness of real humans.

And even a 3-year-old human has better image recognition skills than today’s most powerful computer—which is why Google has licensed the ESP Game and is currently using it with its Google Image Labeler to identify its huge collection of Web pictures.

Self-serving. I wonder what it would cost me to have Professor von Ahn’s ESP Game keyword my collection of 20,000 stock photos? As any stock photographer knows, this is the bottleneck that slows down most image submissions to online agencies.  And even when I get around to doing it, I’m never sure if the keywords I pick are the most relevant!

Limitations. When it comes to technical images and exotic locations, I doubt if I’m going to get quality labels (keywords) using the game approach (I doubt that few 13-year-old kids can correctly identify a ”Whorl Tooth Shark” fossil), so I guess I’m still stuck with doing much of the mundane work myself :-(

Posted by Royce Bair on 06/27 at 10:18 AM

Page 1 of 1 pages