National Geographic News
It turns out you can build Rome in a day—all you need are thousands of photos posted online and a whole lot of computing power.
Since 2005 computer-graphics researchers at the University of Washington in Seattle have been testing whether images collected from the Web can be used to digitally reconstruct buildings in 3-D. (Related pictures: "Ancient Rome Reborn in 3-D.")
The technique could be used to make virtual-reality experiences for tourism, auto-build cities for video games and movies, or help digitally preserve and study historic cities that are being destroyed by human-caused or environmental factors.
The team first developed a program that downloads multiple images of a landmark and finds common points in the architecture. From this information, the program can calculate the subject's three-dimensional structure and where the people who took the shots must have been standing (marked with black cones in the videos above).
But scaling up to build whole cities at once increases the number of source photos needed, upping the time required to download and process them. Searching for "Rome" or "Roma" on the photo-sharing site Flickr, for example, returns more than two million hits, the researchers noted.
From Monuments to Skate Parks
The new version increases photo-matching speed a hundred fold by allowing multiple computers to work in parallel. The update also establishes likely matches before it starts comparing specific features, so that the computers are matching a given picture only to similar images and not to the entire set.
Based on 150,000 publicly accessible Flickr pictures of Rome, the program automatically re-created the Colosseum, Trevi Fountain, and the outside and inside of St. Peter's Basilica, among other Roman icons. With the help of 496 computers, the program completed its monumental task in 21 hours.
Still, you'd need more than 150,000 photographs to re-create Rome in its entirety, team member Sameer Agarwal said in an email. The results are therefore limited to isolated landmarks.
"But interestingly, it doesn't have to be better known [landmarks]," Agarwal added. "We take the bag of images and process them and see what comes out. In some cases, you get interesting things, like a skateboarding park."
The same technique produced a complete 3-D model of the smaller old city of Dubrovnik, Croatia, out of 57,845 pictures in 22.5 hours. Digital re-creations of landmarks in Venice, meanwhile, took 65 hours based on Flickr's collection of 250,000 Venetian snapshots.
However, Agarwal said, "tourist photographs will never capture [a large] city in its entirety. Our hope is that we will ultimately be able to combine the Flickr photographs with something like Google Street View or aerial imagery like Microsoft Virtual Earth to build complete 3-D models of cities."
Initial results to be presented at the upcoming International Conference on Computer Vision in Kyoto, Japan.
SOURCES AND RELATED WEB SITES