Google is asking surfers with time on their hands to help it categorize and label the images indexed by its search engine, building a database of knowledge about the contents of the images.
The company launched a new online game on Friday, Google Image Labeler, which it describes as “a fun way to help us organize all the images on the Web.” In the game, two randomly selected players are each shown the same image, plucked at random from Google’s search index, and given 90 seconds to suggest as many keywords or phrases as they can to describe it. They score points if any of their descriptions match.
Google’s image search engine currently returns results based on captions and other text adjacent to images on Web pages, without reference to the content of the images themselves. The game will allow it to improve the performance of the search engine by returning results based on the players’ descriptions of the images.
The game is not the first attempt to use volunteer labor to create a database of knowledge: The Wikipedia online encyclopedia and the DMoz search directory are two of the better known examples.
Google’s game, based in part on technology developed at Carnegie Mellon University, is not even the first to use volunteer labor to categorize images: The ESP Game developed by Luis von Ahn and other researchers at Carnegie Mellon first put players to work tagging its image database in October 2003.
Building a database using information provided by volunteers has its risks: campaign groups or pranksters might influence or pollute the raw data by associating an insulting term with the image of a political candidate. The campaign to link the term “miserable failure” to the online biography of President George W. Bush is one example of how this can happen.
The ESP Game and Google Image Labeler limit the possibilities for such pranks, since they select players and images at random, and only associate a label with an image if both players independently suggest it.
Von Ahn and his colleagues study “human computation,” or finding novel ways to put human brains to work. Their work includes Peekaboom, a game which harnesses players’ brains to locate objects in images, and Phetch, which goes further than the ESP game by inviting players to create longer descriptions of images. You can also blame them for Captchas, those puzzles featuring sequences of distorted letters that are intended to distinguish between humans and computer impostors.
Google says its search engine indexes billions of images. That may make Google’s goal of labelling all the images on the Web seem far-fetched, especially since players of Von Ahn’s game have only attributed around 17.8 million labels to images since October 2003, according to the game’s Web site.
However, computer users around the world collectively wasted 9 billion hours playing Solitaire on their computers in 2003, Von Ahn estimated in a presentation to Google staff in July. If they had spent that time playing The ESP Game or Google Image Labeller instead, they could have labelled almost 200 billion images.