Oriented Ship Detection via Key Structural Landmarks in Optical Remote Sensing Image

Ship detection from remote sensing imagery constitutes a cornerstone for applications such as marine monitoring and maritime traffic management. However, existing methods based on horizontal bounding box (HBB) or oriented bounding box (OBB) regression paradigms are often plagued by notable limitations in complex scenarios. Traditional HBBs tend to encompass excessive background noise, degrading feature representation efficacy and localization precision. Although OBBs mitigate misalignment, they suffer from angular periodicity discontinuities and high sensitivity to aspect ratios. To overcome these issues, this paper first proposes a novel nine-landmark ship structural representation based on ship functional morphology, which characterizes key regions of the bow, stern, and superstructure as explicit geometric priors. Built upon this representation, we further present a two-stage framework for high-accuracy OBB detection and semantic segmentation. Specifically, an HBB detector first coarsely localizes ship instances. Then, a Ship Holistic-Part Graph Transformer (SHPG-Transformer) module is introduced to explicitly model spatial correlations between global ship geometry and local structural components, thereby improving landmark localization through enhanced structural feature modeling. Furthermore, the Segment Anything Model (SAM) is incorporated to achieve fine-grained semantic segmentation, establishing a unified detection-segmentation pipeline. In the experiments, we extend the HRSC2016 dataset with dense landmark annotations for training and evaluation. Extensive experiments demonstrate that the proposed method achieves state-of-the-art performance on two popular public benchmarks, HRSC2016 and ShipRSImageNet. Moreover, it accomplishes multi-object instance segmentation on HRSC2016, ShipRSImageNet, ISDSD, FGSCR, and FGSD2021 datasets. Project web page: https://vsislab.github.io/OSD-KS/.