Large-scale image geolocalization

Abstract

In this chapter, we explore the task of global image geolocalization–estimating where on the Earth a photograph was captured. We examine variants of the õim2gpsö algorithm using millions of õgeotaggedö Internet photographs as training data. We first discuss a simple to understand nearest–neighbor baseline. Next, we introduce a lazy-learning approach with more sophisticated features that doubles the performance of the original õim2gpsö algorithm. Beyond quantifying geolocalization accuracy, we also analyze (a) how the nonuniform distribution of training data impacts the algorithm (b) how performance compares to baselines such as random guessing and land–cover recognition and (c) whether geolocalization is simply landmark or õinstance levelö recognition at a large scale. We also show that geolocation estimates can provide the basis for image understanding tasks such as population density estimation or land cover estimation. This work was originally described, in part, in õim2gpsö [9] which was the first attempt at global geolocalization using Internet–derived training data.