First place in the ACM Multimedia 2023 UAVs in Multimedia Challenge

2 Oktober 2023

Unmanned vehicles, especially drones, require precise geo-localisation to transport entities from one location to another. The widespread use of drones has made high-quality aerial footage accessible to a broader audience, presenting an opportunity to integrate this footage for GPS-less geo-localisation or location correction. In this context, the "UAVs in Multimedia: Capturing the World from a New Perspective" workshop at ACM Multimedia emphasized UAV navigation and coordination using artificial intelligence techniques. A notable highlight was a challenge based on the University-160k dataset to test visual geo-localisation methods. Participants were presented multiple UAV-captured views of buildings. The task was to match these views with referenced satellite images for precise location pinpointing.

The winning approach, presented by Fabian Deuser, Konrad Habel, Norbert Oswald from UniBwM, and Martin Werner from TUM, introduced an orientation-guided training framework for UAV-view geo-localisation. This method estimates UAV image orientations relative to satellite imagery using a lightweight prediction module based on contrastive learned embeddings. Our approach outperforms previous methods and achieved top results on the University-160k challenge. The paper will be presented at the ACM MM 23 in Ottawa, Canada.

Orientation-Guided Contrastive Learning for UAV-View Geo-Localisation
Fabian Deuser, Konrad Habel, Norbert Oswald

[arXiv], [ACM MM 2023], [Challenge Homepage]