A very interesting idea and implementation were proposed in [8]. The authors utilize deep networks to geolocalize a photo by directly learning to match street-level images to aerial images without using ground-level reference imagery. This work is close to our approach but evaluating on quite different and specific input data.