3 D Exploitation of 2 D Imagery

The invention of the digital camera is attributed to Eastman Kodak engineer Steven Sasson in 1975 [1]. Sasson’s device recorded 0.01-megapixel black-and-white pictures to cassette tapes. The first digital photograph required 23 seconds to generate and needed a television set to display [2]. Today, digital cameras recording greater than 10-megapixel color photos to Secure Digital (SD) cards with multi-gigabyte storage capacities are commonplace. Though the total number of electronic images in existence is not known, billions of photos and video clips are now accessible on the World Wide Web. The current capability to collect and store digital images vastly outpaces the current capability to mine digital pictures. Existing archives of photos and videos are basically unstructured. As anyone who has ever tried to find some particular view of interest on the Internet knows, querying imagery websites can be a frustrating experience. Text-based searches generally do not return salient metadata, such as camera geolocation, scene identity, or target characteristics. Some basic organizing principle is consequently needed to enable efficient navigating and mining of vast digital imagery repositories. Fortunately, three-dimensional (3D) geometry provides such an organizing principle for imagery collected at different times, places, and perspectives. For example, a set of photos of some ground target represents two-dimensional (2D) projections of 3D world space onto a variety of image planes. If the target’s geometry is captured in a 3D map, it can be used to mathematically relate different ground photos of the target to each other. Moreover, as the diagram in Figure 1 indicates, the 3D map connects Recent advances in computer vision have enabled the automatic recovery of camera and scene geometry from large collections of photographs and videos. Such three-dimensional imagery reconstructions may be georegistered with maps based upon ladar, geographic information system, and/or GPS data. Once 3D frameworks for analyzing two-dimensional digital pictures are established, high-level knowledge readily propagates among data products collected at different times, places, and perspectives.