A 3D scene from strangers’ photographs
At SIGGRAPH 2006 Noah Snavely, Steven Seitz (University of Washington) and Richard Szeliski (Microsoft Research) presented Photo Tourism. The system worked out where each of thousands of random internet photos of a landmark had been taken from, built a sparse 3D model from them, and let one move through it from shot to shot. For Notre Dame, 597 of 2635 Flickr photographs were registered.
Why it matters
Structure from motion had until then been tested on far simpler sets taken under uniform conditions. The authors call their paper its first successful demonstration on the kind of photographs found on Google and Flickr: hundreds of different cameras, seasons and lighting. A collection of strangers’ snapshots became a measurement of a place.
ACM Transactions on Graphics 25(3), pp. 835-846. SIFT finds the features, pairs of photos are matched, and cameras and points are refined by iterative bundle adjustment. The sets: Notre Dame, 597 photos registered from 2635 found on Flickr with the query "notredame AND paris"; Trevi Fountain, 360 of 466; Half Dome in Yosemite, 325 of 1882; Great Wall, 82 of 120; Prague, 197 of the authors’ own. Reconstruction took from a few hours for the Great Wall to about two weeks for Notre Dame. An annotation added to one photo passes by itself to every other photo in which the same object is visible. The record does not link the paper to Photosynth: the word is not in it, and no separate source on the Microsoft product was read.