Visual Localization

Image geo-localization is the task of estimating the location of a query image using a set of geo-tagged reference images. This paper presents a new approach based on geospatial mean shift for image geo-localization. In the proposed approach, the nearest neighbor for each SIFT feature point in the query image is first identified from the reference set. Next, the region with the highest density of feature matches is found using the mean shift method, because each scene is typically visible, at least partially, in multiple reference images captured in its vicinity. We propose a density function which is consistent with the location distribution. Finally, a geometric verification step is applied in a one-to-one matching fashion. The effectiveness of the proposed approach is demonstrated on a city-scale dataset of street view images.