Researchers have developed a way to build 3D images based on similar photos.
Using millions of images pulled from Flickr, a team from the University of North Carolina at Chapel Hill (UNC) were able to build 3D models of historical landmarks, and even the majority of an entire city.
The algorithm developed by the UNC team isn’t perfect, but it can get a lot done within 24 hours. It starts by grouping similar-looking photos into “appearance clusters”. Only groups with any geometric compatibility—clusters containing photos of the building’s front and sides for instance—are retained. These remaining clusters are used to build the 3D models themselves, with the source photos manipulated to “paint” the structure determined by the algorithm.
The detail of the resulting models is quite surprising, though it’s clear that the UNC team has a long way to go. Some of team’s models are incomplete, while commissioning 3D modelers would still produce more accurate (and expensive) results. Still, the project is yet another example of crowd-sourcing’s advantages.