IceHorizon dataset: hybrid AI beats classical methods for polar horizon detection

Horizon detection is critical for autonomous navigation in polar regions, but ice-covered waters make it uniquely difficult: low water-sky contrast, cluttered ice structures, and variable lighting throw off traditional algorithms. To address this, researchers Alisa Pesotskaia and Emin Zerman introduced IceHorizon, a new benchmark dataset containing 30 ship-based and 8 drone-based videos. They systematically compared six horizon-detection algorithms—four classical computer vision methods (such as edge- and color-based techniques) and two hybrid approaches that combine deep learning segmentation with classical line detection like Hough transforms.

Results show hybrid methods consistently outperformed purely classical ones, delivering the most accurate and stable horizon estimates, especially in visually ambiguous scenes. Performance was notably higher on ship-based imagery than drone-based footage, a gap the authors attribute to differences in camera motion, field of view, and image quality. The findings suggest that relying on classical CV alone is risky for real-world icy navigation, while hybrid pipelines offer a more robust path. IceHorizon and all code are publicly available, enabling further research into trustworthy maritime AI systems.