A Systematic Evaluation of Image Geolocation Inference Models

Lecture Center Concourse

The ability to determine the location of a photo based on its content has become a major focus amongst researchers today, which can enable location-based services without requiring GPS data and may raise privacy concerns at the same time. Photo geolocation refers to the exact location coordinates where the photo was taken. Many methods and models have been proposed in an attempt to unveil a flawless photo geolocation inference technique. For example, Im2GPS is a technique that matches a query image against a database of 6.5 million Flickr images using global image descriptors and returns the location of the most closely related image. PlaNet is a convolutional neural network (CNN) that generates a probability distribution over the earth and assigns a likelihood that the queried photo was taken within a given region. This project seeks to present and systematically evaluate the various methods that have been proposed in one comprehensive paper. The proposed techniques are compared in their methodology as well as their results in order to provide thorough insight into the development and success of photo geolocation inferencing. The methodology of the techniques considered most effective would then provide a direction for future research to focus and build upon. I will also identify limitations and open challenges for photo geolocation inferencing.