First place in the ACM Multimedia 2024 UAVs in Multimedia Challenge

Unmanned Aerial Vehicles (UAVs) rely on accurate geo-location for navigation, but traditional GPS-based methods can fail in urban environments, in bad weather, or due to signal interference. To address this, AI-based visual geo-localization matches UAV-captured imagery to satellite references, enabling accurate positioning without GPS. However, real-world challenges such as lighting changes, occlusions, and weather variations make this task complex.

The UAVs in Multimedia: Capturing the World from a New Perspective workshop at ACM Multimedia continues to advance UAV localization research. At the 2024 competition in Melbourne, Australia, organizers presented the University-160kWX dataset, designed to test robustness under varying weather conditions. The dataset contains 160,000 satellite images and over 80,000 drone-view images. Each drone view, captured under different weather scenarios, must be matched with its corresponding satellite image for localization. To increase the complexity of the task, the dataset contains a large number of distractor satellite images without matching drone views. Our team - Fabian Deuser, Konrad Habel, Norbert Oswald (UniBwM), and Martin Werner (TUM) - secured first place again with an improved re-ranking approach to handle uncertain predictions. Using Vision Transformers pre-trained with DINOv2 and a novel k-Means re-ranking strategy, we achieved state-of-the-art accuracy in visual localization.

For further insights, refer to our attached paper, which details our methodology and findings. Our work was presented at the UAVM Workshop of the ACM Multimedia 2024 in Melbourne.

Optimizing Geo-Localization with k-Means Re-Ranking in Challenging Weather Conditions

Fabian Deuser, Konrad Habel, Martin Werner, Norbert Oswald

[LINK]