Crowdsourcing the World's Landmarks

Chris Dixon has a really interesting post today. In it, he talks about having the foresight to structure information in order to get better and better performance from what he calls "collective knowledge systems." Toward this end, Chris thinks that academic computer scientists could play an important role. "Unfortunately," Chris writes, "academic computer scientists still seem to model their field after the 'hard sciences' instead of what they should model it after -- social sciences like economics or sociology."

Well, as life would have it, I read Chris's piece while in the midst of collecting links for my own post about research being done at Cornell University, concerning the organization of millions of geo-tagged photos on Flickr and how best to mine and map that "collective knowledge."

Here's a link to a paper describing that research, "Mapping the World's Photos," written by David Crandall, Lars Backstrom, Dan Huttenlocher (who is dean of Cornell's Faculty of Computing and Information Science) and Jon Kleinberg. They point out that photo-sharing sites have connective structure, provided by the millions of people that take and post the photos. This is significant because:

". . . it means that we can discover, through collective behavior, what people consider to be the most significant landmarks both in the world and within specific cities; which cities are the most photographed; which cities have the highest and lowest proportions of attention-drawings landmarks; which views of those landmarks and most characteristic; and how people move through cities and regions as they visit different locations within them."

Check out these maps generated by the Cornell researchers -- or, I should say, maps that are generated from the programs they built to "automatically [mine] the information latent in large sets of images."

And here is a detail from one of those maps, the one for Paris:

I offer this detail because I can relay a story Dean Huttenlocher shared with Cornell alumni at an event last week in Bellevue: the photo of the Eiffel Tower taken at the bottom, just to the left of the Musée d'Orsay, represents the canonical view of the landmark for American tourists; whereas the image labelled "Tour Eiffel," just below the image of the tower shot from Trocadéro, represents the canonical view as determined by residents of France. Apparently, French people take as given that one must always cross the river to get a proper picture of the tower. (My daughter and I are debating this evening what this might mean, culturally.)

How do the Cornelleans determine which image is the most representative?

"[W]e pose canonical image selection as a graph problem. We construct a graph in which each node represents a photo and between each pair of nodes is an edge with the weight indicating the degree of visual similarity between the two photos. Our goal is then to find a tightly-connected cluster of photos that are highly similar. To do this we use a spectral clustering technique that partitions the nodes of the graph using the second eigenvector of the graph's Laplacian matrix. Finally, we choose as the canonical image for each cluster the one corresponding to the node with the largest weighted degree."

Okay, so I don't know what that means. But I gotta give Daniel Carleton, Bruce Roberts and certain of my other readers something more to chew on from time to time! (And besides, the words have dimension and weight, like silver coins in your palm.)

Coming back full circle: I can't dispute Chris's point about university computer science departments, in general. At Dean Huttenlocher's talk last week, he also suggested that change comes slowly to academic disciplines. But at Cornell, the Computer Science Department "co-sponsor[s] a major in Information Science, Systems, and Technology in Engineering and a major in Information Science in Arts & Sciences and Agriculture & Life Sciences." You can get a degree in anything at that school, practically.