Approaching visual geo-localization through classification

The increasing interest towards visual geolocation (or visual place recognition) has been noticeable in recent years. A fast and efficient system able to identify a particular place around the globe, that uses only the visual content of a query image, has been requested in many different fields, going from virtual reality to self driving cars and exploratory robots. Advances in artificial intelligence and dense open source datasets boosted the research community on proposing a set of alternative ways to tackle this challenge, being retrieval and classification two of the most diffused approaches, each of which propose its own advantages and weaknesses. Although, there is no absolute winner when comparing the existing approaches, more efficient systems can be obtained when combining more than one technique through the processing pipeline. The objectives of this project were focused on the analysis of the classification approach. Primarily consisting of partitioning the geographical area of interest into geographical cells and developing machine learning models able to associate a particular set of images to its correspondent geo-class. The goal of this project was to study, implement and deploy existing approaches, popular among the existing literature, but whose authors in some cases never released the source codes to replicate their results. A second part of the project consisted of developing a deep, complete and updated comparison of the existing methods, studying their performances on highly dense datasets of urban scenarios.