Large-scale geolocalization of overhead imagery

Abstract

In this paper, we investigate state-of-the-art computer vision techniques to perform large scale geolocalization of overhead imagery through image matching. We consider two types of features: scale invariant feature transform and region-based shape features. Since these features can be high dimensional and an image can contain many of them, using them to perform image matching can be computationally expensive. Therefore, we also investigate two methods for performing efficient matching: aggregating the features at the image level using a bag of words framework and using hashing to perform multiple, efficient matches and then aggregating the results. We show that hashing performs better in terms of accuracy but is expensive computationally compared to bag of words. We also show that shape features may be accurate and efficient for small data sets, but they do not scale well to large data sets.