September 18, 2006
This just in!
Anyone who has kids knows sometimes you have to make learning seem more like playing to keep your kids’ interests.
You click to find a partner who pairs up with you. You are given 90 seconds to label images. You and your partner see the same images and you must label them…each time you label them with the same keyword, you are given 100 points and a new image is shown. There are terms that are offlimits…perhaps words they already associate with the picture because others have already labeled them or they have already extracted those terms and they are somehow verified/validated (sort of a way to prevent duplicate work)?
The idea behind letting 2 people do it instead of one is 1) it becomes a game and could create some sort of social aspect to it and 2) each person’s answer helps validate each others. Pairing them up in realtime as opposed to using someone’s answer from a previous “test” just makes it seem more like a game (kind of like where you log in to Yahoo Games and wait for a challenger to sign on) but it also helps push the timed test thing since you know that someone is depending on you to get some points (a team mentality).
Its a nice way for Google to get “free” help from its users to help improve search while keeping their interest in doing it. Google doesn’t hide this fact either:
Welcome to Google Image Labeler, a new feature of Google Image Search that allows you to label random images to help improve the quality of Google’s image search results.
What is required to participate?
Just an interest in helping Google improve the relevance of image search results for users like yourself. Although you do not have to log in to your Google account to help, logging in will allow you to keep track of your points. You can also choose to provide a nickname, or you can remain anonymous.
I wonder if they will ever do a Flickr-type of feature with this…
Creating effective algorithms for image search has always been a challenge. You can try seeing how many people link to an image (or the page an image is on) with certain key phrases in the links, you can check alt tags, you can check title tags within an href tag linking the image, you can check surrounding metadata (other data/keywords on the same page as an image), you can read an image’s metadata…but ultimately, you are unable to “see” the image algorithmically. I’ve tried OCR on images with limited success as well as scanning or certain colors and hues to try to help identify images…both with limited success and are also computationally expensive. I’ve even seen other types of image searches such as search by sketch, person (facial recognition), colors, and others but they are also limited for various reasons. It is also the reason why captchas are effective.
Google’s Image labeler, however, gives you something to help improve these inherent problems…a cheap (free) editorial staff that has a community of “validators” and synonym creators.