Be part of a tag team... - Net.Matters - June 2008

A problem that many users of the web experience is trying to source the right image.  This can be down to a number of reasons: poor tagging i.e. labelling of the images and irrelevant descriptions.  This means that search engines won't return these images on their search results page - and we don't see them. 

A professor in computer science from Carnegie Mellon University in Pittsburgh has been addressing the issue of poor image labelling for the last 5 years.  He first developed a game back in 2003 (Extra Sensory Perception Game) which pits 2 players against each other to label an image that both can see.  Each player has to guess what their opponent has written to describe that image.  If they agree, the description is used as an explanatory tag.  The same image is used in subsequent games with other players which enables the build up of a detailed label for that image.

Reports so far suggest that as many as 100m images have been labelled using this method and even Google has got involved by buying a licence from Professor von Ahn to bring out its own version of the game.  Google's game is entitled Google Image Labeler and was released in 2006 in order to help the quality of the results that are returned.

According to Professor von Ahn his game could be the answer to all those poorly labelled images on the web - he believes that they could all be labelled within weeks, if the game appealed enough to would-be players.  As many as 1m players have registered to play but only around 200,000 of those registered actually play the game regularly.  To try and bring in more players, Professor von Ahn's team has introduced other games to see if they have more appeal.  The site to visit is www.gwap.com.  One of the games is called Squigl and players are prompted to use their mouse and trace around an image.  Points are won by those who closely match the other player's tracing. 

"This really trains the computer to see.  If we outline enough pictures of dogs, for example, the hope is that in the long-run the computer will know itself the outline of a dog."  Professor von Ahn explains.

Another game on the site is called Matchin.  The player is asked to select from 2 images which one he thinks is the most attractive.  The aim of this is to get a database of the most appealing images on the internet.  For more information on these image labelling games visit, www.gwap.com or Google's version, http://images.google.com/imagelabeler/.