Abstract
Accurate semantic boundary detection in maritime environments is critical for enabling robust autonomous navigation, particularly in GPS-denied scenarios. These environments pose significant challenges due to water reflections, adverse weather conditions, and obscured vision, which complicate boundary estimation tasks and visible horizon line detection. This paper presents a novel approach leveraging a curriculum learning scheme to address the difficulties of sparse labelling of fine semantic boundaries. Central to our framework is the introduction of a novel loss function, the Fourier Spectral Alignment, which models semantic boundaries as the Fourier series. Positioning the loss with spatial and frequency components enables the model to precisely capture complex boundary characteristics by aligning spatial and frequency domain representations. The proposed method demonstrates superior performance in low-data regimes, achieving high accuracy in boundary detection even under challenging environmental conditions. Experimental results on a dataset specific to the maritime domain highlight the robustness and efficacy of our approach. We improve upon existing approaches across all metrics, setting a new standard for semantic boundary detection and semantic segmentation in the maritime domain.