Diving Deeper into IM 2 GPS

Predicting the location an image was taken at is a long standing research problem in computer vision. Given the recent successes with using deep learning techniques, we present a series of approaches towards image geo-location using a variety of deep learning techniques. This report describes our implementation and experimental set-up of training the GoogLeNet and AlexNet architectures on a sample sub-section of the im2gps dataset. Our results achieve competitive performances as compared to the current state of the art. Thus, our major contribution for this report is a methedological bridge between the existing frameworks fro image geo-location. Keywords—computer vision, classification, location, deep learning.