Urban Image Geo-Localization Using Open Data on Public Spaces

In this paper, we study the problem of urban image geo-localization, where the aim is to estimate the real-world location in which an image was taken. Among the previous approaches to this task, we note three distinct categories: one only analyzes metadata; the other only analyzes the image content; and the third combines the two. However, most previous approaches require large annotated collections of images or their metadata. Instead of relying on large collections of images, we propose to use publicly available geographical (GIS) data, which contains information about urban objects in public spaces, as a backbone database to query images against. We argue that images can be effectively represented by the objects they contain, and that the spatial geometry of a scene - i.e., the positioning of these objects relative to each other - can function as a unique identifier for a particular physical location. Our experiments demonstrate the potential of using open GIS data for precise image geolocation estimation and serve as a baseline for future research in multimedia geo-localization.