Towards better data variation in maritime perception datasets

Ensuring the safety and reliability of AI-based perception systems for autonomous maritime navigation critically depends on validating these systems under realistic and diverse conditions. However, the maritime sector currently lacks standardized, comprehensive datasets that capture the variability of real-world operational and environmental scenarios. To address this gap, we propose a novel approach for systematically evaluating the visual quality of maritime image datasets using a Gabor filter-based metric. This method assesses the structural information content in images and effectively distinguishes between high and low visibility conditions, which is essential for testing perception systems’ robustness. We demonstrate the approach on a large dataset collected aboard a ferry operating in the Øresund, highlighting its ability to automatically score and rank images by visibility. Our results show that the proposed metric facilitates the creation of more representative and challenging datasets, thereby supporting the development and validation of safer and more reliable autonomous maritime perception systems.