Enhancing Worldwide Image Geolocation by Ensembling Satellite-Based Ground-Level Attribute Predictors

We examine the challenge of estimating the location of a single ground-level image in the absence of GPS or other location metadata. Geolocation systems are typically evaluated by measuring the Great Circle Distance between a single predicted location and the ground truth. Because this measurement only uses a single point, it cannot assess the quality of an estimated set of regions or score heatmaps. It is critical to characterize the distribution of potential geolocation areas to help when a system is applied to less well-sampled regions, such as rural and wilderness areas where finding the exact location may not be possible. This evaluation may help characterize a system in relation to follow-on procedures that further narrow down or verify predicted locations. This paper introduces a novel metric, Recall vs Area (RvA), which assesses distributions of location estimates. RvA treats image geolocation results in a manner similar to precision-recall in document retrieval, measuring recall as a function of area. For an ordered list of (possibly non-contiguous) predicted regions, we measure the accumulated area required for the region to contain the ground truth coordinate. This produces a curve analogous to precision-recall, enabling evaluation for varying search budgets. This view of the problem inspires a simple ensembling approach to global-scale image geolocation that combines multiple models, attribute predictors, and data sources. Specifically, we combine the geolocation models GeoEstimation [11] and GeoCLIP [2] with attribute predictors based on ORNL LandScan [16] and ESA Land Cover [1]. We find notable improvements in non-urban and under-represented areas on Im2GPS3k and Street View datasets.