IceHorizon: A Dataset for Horizon Detection in Ice-Covered Maritime Environments and Comparative Evaluation of Detection Methods

Imagine you are the captain of a ship, or perhaps a pilot flying a drone, trying to navigate through a world that looks like a giant, frozen mirror. In this icy realm, the sky and the ocean aren't two different colors; they are both a blinding, shimmering white. Finding the "horizon"—that invisible line where the sky meets the sea—is like trying to spot a ghost in a snowstorm. If you can't find that line, your ship might tilt dangerously, your camera might get dizzy, and your navigation system could get lost. This is the tricky puzzle of "horizon detection" in icy waters. Scientists use special computer tricks to solve this: some are like old-school detectives looking for sharp edges and lines, while others are like modern artists who have learned to "see" the difference between water and sky by studying thousands of pictures. But until now, most of these computer tricks were trained on sunny, open oceans, leaving them confused and clumsy when faced with the messy, reflective chaos of ice.

This paper, titled "IceHorizon," steps into that frozen confusion to see which computer detective is the best at finding the horizon in the ice. The researchers built a brand-new library of 38 videos, filmed from both ships and drones, showing exactly what it looks like to sail through icy waters. They then put six different computer methods to the test: four that use classic, rule-based math (the old-school detectives) and two that mix deep learning with those rules (the hybrid artists). They measured how close the computer's guess was to the real horizon, how much of the horizon it could see, and how fast it worked.

The results were a clear victory for the "hybrid" methods. When the computer tried to find the horizon on the ship videos, the hybrid approaches were up to 3.5 times more accurate than the best of the old-school methods. Even on the drone videos, which were much harder to figure out, the hybrids were 1.6 times better. The old-school methods, while very fast, often got lost in the glare of the ice or the reflection of the waves, sometimes missing the horizon entirely or drawing it in the wrong place. The hybrid methods, however, were like having a guide who knows what ice looks like; they could ignore the confusing reflections and find the true line. Interestingly, every single method worked better on the ship videos than on the drone videos. The drone footage was just too messy, with fog, mountains, and weirdly curved horizons that confused even the smartest algorithms. The paper concludes that while we still need more data to make these systems perfect for high-altitude drone flights, mixing deep learning with classic line-finding is currently the most reliable way to keep our ships and drones steady in the frozen north.