By Roland Piquepaille
Computer scientists from Carnegie Mellon University (CMU) are developing and using online games to train computers to better see according to the Pittsburgh Post-Gazette. One of these games is Peekaboom, used online for free by teams of two players. The first one, designated as "Peek," sees on his screen an image -- initially empty -- and a word that describes the image or one element of the image. The second one, named "Boom," gradually reveals the image or gives hints to "Peek" until he correctly guesses the word associated to the image. And this use of "segmented" images might improve "computer vision" by "teaching" them how to identify objects. But read more before playing...Below is a simplified trial run of Peekaboom, where "Boom" is the Revealer and "Peek" the guesser (Credit: Peekaboom.org).
You'll find a larger version of this image -- along with other details -- on page 15 of this presentation named "Peekaboom: A Game for Locating Objects in Images" (Powerpoint format, 23 pages, 809 KB).
The images and words used in Peekaboom, created by Luis von Ahn, have been taken from a previous CMU online game, the ESP Game, in which participants generate descriptive labels for images.
Now, let's read some short excerpts from the Pittsburgh Post-Gazette article, starting with the game itself.
In the game, which graduate student Roy Liu has been programming since September, one player is given an image and a word that describes the image or one element of the image. The object of the game is for the first player, designated as Boom, to get the other player, designated as Peek, to correctly guess the word.
Peek initially sees only a blank screen, but as play proceeds Boom reveals the image, little by little, by moving his cursor over the image.
When the word is correctly guessed, or when the players agree to pass on an image, the players switch roles and play resumes. The players receive scores based on the number of correct guesses they can make in four minutes.
But how can this game improve computer vision?
In the early days of machine vision research, it was assumed that computers could learn to identify an object, such as a car, or a spoon, or a face, if it was given the rough geometry of the object, said Alexei Efros, a computer vision researcher at Carnegie Mellon. But that wasn't a successful approach.
A better approach to teaching a computer how to identify an object, such as a car, is to show it lots of images of cars, of various makes and colors, taken from a variety of angles and distances and under a variety of lighting conditions, he said.
For more information, you can check how Peekaboom works by reading this introduction (PDF format, 1 page, 461 KB).
But for fun, just play with Peekaboom!!!
Sources: Byron Spice, Pittsburgh Post-Gazette, August 1, 2005; and various web sites
Related stories can be found in the following categories.
And remember that comments are no longer accepted here. If you want to tell me something about this post, please go to the bottom right of this page and send me an e-mail.
Famous quotes containing the word teach:
“We must continually remind students in the classroom that expression of different opinions and dissenting ideas affirms the intellectual process. We should forcefully explain that our role is not to teach them to think as we do but rather to teach them, by example, the importance of taking a stance that is rooted in rigorous engagement with the full range of ideas about a topic.”
—bell hooks (b. 1955)