Improving semantic orthophotos by a fast method based on harmonic inpainting

Context-based modeling of 3D urban terrain has become increasingly popular in the last two decades. Typically, orthophotos are used for texturing ground. In order to increase locational awareness, it is useful to eliminate from the orthophotos those object instances which frequently appear and disappear in the terrain. Vehicles are good examples of such instances. Assuming that vehicles were detected at a previous stage of the algorithm, we developed a simple, easily parallelizable procedure on orthophoto inpainting while considering classes of the surrounding objects. This is done by reducing the weights of pixels which should not be used for inpainting. The benefits of our approach are demonstrated for two datasets by comparison with a state-of-the-art method and by integrating the corrected orthophoto into the already existing urban terrain model which moreover includes the ground surface as well as 3D models of buildings, trees and vehicles.