Computer Vision News - July 2021

Presentation 26 Best of CVPR 2021 Nowadays, huge databases of satellite images from every part of the world are widely available via services like Google Maps . However, the coverage of street- level imagery is more variable. In this paper, the team take street-view images from unknown locations and match them against a GPS-tagged satellite database to find the closest satellite image. As humans, we are more used to seeing street-view images than we are images taken from above, so it may not be immediately obvious to us when comparing a satellite image to a street-view image whether they are showing the same location. This is made even more difficult because they are likely to be taken at different times of the day by different cameras. That is where computer vision and deep learning can help. Aysim Toker is a PhD student at the Munich School for Data Science. Her first supervisor is Professor Laura Leal-Taixé and her second supervisor is Professor Xiao Xiang Zhu from the German Aerospace Center and the Technical University of Munich. Her paper explores a novel geo-localization method that may just revolutionize our everyday GPS. She speaks to us ahead of her presentation today. Coming Down to Earth: Satellite-to-Street View Synthesis for Geo-Localization