Leveraging Topographic Maps for Image to Terrain Alignment

With the wide-spread availability of photographic and cartographic data, it becomes desirable to be able to geo-localize any picture in the world. Existing approaches have so far shown impressive results, but they are still lacking in either precision or applicability. In the present work, we explore as an additional cue, semantic image labeling coupled with topographic maps. As an intermediate step towards the ultimate goal of universal geo-localiztion, we show that these cues are suitable for estimating the viewing direction of a terrestrial image, given the image's location.