We are proud to announce that a team led by Fabian Deuser won the first place at the cross-view geolocalization competition.
In this challenge, the organizers presented a novel, challenging cross-view geo-localization dataset, called University160k. The motivation was to provide a comparably large satellite-view dataset for geolocalization to increase the number of similar features in different images. With a strategy of using pseudolabels to get a good alignment of latent space features and the localization problem, we were able to outperform all other submissions in the challenge. A paper on this topic is accepted and will be presented in the workshop.
Congratulations to our fresh (first month?) PhD student Fabian Deuser, who was the lead in all of this work.
This challenge has triggered a new line of research in our group as we believe that the true geolocalization problem is even harder than already depicted in the enlarged University160k dataset. On the other hand, the localization problem is typically a local problem as a coarse location might already be known in most applications. Furthermore, we will extend this activity to the indoor space, where even a limited-scalability reliable indoor geolocalization from images would be very helpful.
So stay tuned for the workshop presentation, the paper, and our follow-up work maybe including additional geolocalization challenges.