In computer vision, the photo geolocation problem has been usually approached at a global scale or regional

scale. In this project, we derive knowledge from urban studies and present a classification based method looking into neighborhood scale photo geotagging.

We subdivide Boston Chinatown into multiple geographic cells and train a deep network using 20K google street view images labeled with local restaurant density and street speed limit. We show that this multitask model could output a probability distribution over a couple of cells in the neighborhood.