DeepGeo: Photo Localization with Deep Neural Network
This addresses the challenge of photo localization for applications in computer vision and learning, though it is incremental as it builds on existing deep learning methods with a new dataset.
The paper tackles the problem of predicting the geographical location of an image by framing it as a classification task across 50 U.S. states, achieving 71.87% accuracy with top-5 guesses and outperforming humans in GeoGuessr.
In this paper we address the task of determining the geographical location of an image, a pertinent problem in learning and computer vision. This research was inspired from playing GeoGuessr, a game that tests a humans' ability to localize themselves using just images of their surroundings. In particular, we wish to investigate how geographical, ecological and man-made features generalize for random location prediction. This is framed as a classification problem: given images sampled from the USA, the most-probable state among 50 is predicted. Previous work uses models extensively trained on large, unfiltered online datasets that are primed towards specific locations. To this end, we create (and open-source) the 50States10K dataset - with 0.5 million Google Street View images of the country. A deep neural network based on the ResNet architecture is trained, and four different strategies of incorporating low-level cardinality information are presented. This model achieves an accuracy 20 times better than chance on a test dataset, which rises to 71.87% when taking the best of top-5 guesses. The network also beats human subjects in 4 out of 5 rounds of GeoGuessr.