Eyespy: Supporting Navigation through Play

Comment on Bill's blog.

Eyespy is an interesting program built by Marek Bell, Stuart Reeves, Barry Brown, Scott Sherwood, Donny McMillan, John Ferguson and Matthew Chalmers that uses a game to get humans to build a database of pictures that are useful for navigating an area. How is this accomplished? The game it self is divided into two parts. First, the player goes around (literally anywhere) and creates photo or text tags in places for other players to find. However, these need to be easily identifiable since when another player 'confirms' your tag, you get extra points. The second part of the game is to find tags created by other players. Each day you are sent five photo tags and five text tags. By finding these you and the player that created the tag earn points.

The tag and confirm portions must be done in approximately the same locations. For example, if I recognize where a picture was taken, I can't just confirm it from my house (unless the picture was taken at my house). I have to go to the place where the picture was taken. The way this works is that the phone you are playing on detects the nearby wireless networks to determine your location.

The makers of Eyespy then conducted a user study with images captured from Eyespy and images pulled from Flickr. Two routes were created through a large public area and users used photos pulled from either Eyespy or Flickr exclusively to try and find their way through the route. The users that used the Eyespy tags were able to identify the locations on the route much faster and with much greater accuracy than those that used the Flickr tags. This could be used to help aid tourists in navigating unfamiliar areas.

I think this is a great way to improve location tags on images, especially if the Eyespy becomes more widespread. It is a very reliable way for getting identifiable images since the ones that are most identifiable will most likely get the highest scores. And the incentive of high scores encourages players to try and snap easily identifiable pictures. I would love to see if this 'human algorithm' (as the authors call it) could be applied to other areas of computer science.