Danny Veikherman (EE, Technion)
Light-field imaging can be scaled up to a very large area, to map the Earth's atmosphere in 3D. Multiview spaceborne instruments suffer low spatio-temporal-angular resolution, and are very expensive and unscalable. We develop sky light-field imaging, by a wide, scalable network of wide-angle cameras looking upwards, which upload their data to the cloud. This new type of imaging-system poses new computational vision and photography problems, some of which generalize prior monocular tasks. These include radiometric self-calibration across a network, overcoming flare by a network, and background estimation. On the other hand, network redundancy offers solutions to these problems, which we derive. Based on such solutions, the light-field network enables unprecedented ways to measure nature. We demonstrate this experimentally by 3D recovery of clouds, in high spatio-temporal resolution. It is achieved by space carving of the volumetric distribution of semi-transparent clouds. Such sensing can complement satellite imagery, be useful to meteorology, make aerosol tomography realizable, and give new, powerful tools to atmospheric and avian wildlife scientists.