An international team of researchers has developed a new way to turn photographs from the media-sharing sites like Flickr into intricate 3-D computer models using only a home computer.
The new technique—developed by researchers at the University of North Carolina at Chapel Hill and ETH Zurich—streamlines a complex process that previously required clusters of dozens of computers. Researchers say they believe the new technology could spark a host of 3-D-modeling projects that are crowdsourced, with photos taken by tourists and posted online. “The data acquisition and computation can be outsourced to the people themselves,” said Jan-Michael Frahm, an assistant professor of computer science at Chapel Hill and a lead researcher.
To show off the technique, the research team created 3-D models of several famous landmarks in Rome (the same challenged tackled in a similar visualization project at the University of Washington). Researchers in North Carolina used a complex algorithm to scour the massive Flickr database and collect more than three million user-generated images of Rome’s most famous landmarks, including the Colosseum and Trevi Fountain. The algorithm then looked at overlapping images to find common reference points in the photographs—a sharp edge on the Trevi Fountain, for example—and used that to determine where the photographers were standing relative to the object when they snapped their pictures. That information was then used to create a realistic 3-D rendering of the object’s physical features, seen from many angles.
Although building 3-D models from 2-D-image databases is not new, the researchers say they did break new ground in sorting through database images. Past algorithms have made use of all images in the database; in this case, the researchers used an image-analysis tool to group similar photos and use representative images instead of working with thousands of duplicates. That approach is especially helpful at landmarks where herds of tourists often snap photos from the same spot. “We don’t compare three million images to three million images,” said Marc Pollefeys, a professor of computer science at ETH Zurich and a collaborator on the project. “We can have a lot less work and have a model that’s essentially the same.”
The final models are similar to 3-D maps seen on Google Earth, which Mr. Frahm said does much of its modeling manually. Because researchers use an algorithm that generates the models automatically, Mr. Frahm said their technique can be used to create much more detailed models and, through crowdsourcing, can provide virtual access to more places. With Google Earth, “you can only go where Google cameras go,” Mr. Frahm said. “Here, we’re loosening up the restrictions on where you can go.”
The researchers have also created 3-D models of famous landmarks in Berlin, and they are working on expanding coverage worldwide.
They hope that the technology will one day make it possible for everyone to create their own 3-D models, whether it’s gamers recreating their homes in a Sims-like virtual world or the casual tourist creating a 3-D model to remember a Roman holiday. “You download your pictures at home and essentially recreate your favorite sites in the city,” Mr. Pollefeys said. “You can personalize it.”