Hierarchical YOLO26 with Axis-Wise Context Block and Poincaré Supervision for Fine-Grained Ship Detection in Remote Sensing

Fine-grained oriented ship detection remains challenging because visually similar vessel categories may exhibit substantial variations in scale, aspect ratio, and orientation. To address these challenges, this study proposes YOLO26-OBB-HMS (Hierarchical Multi-Label Supervision), a hierarchy-aware extension of YOLO26-OBB that integrates a pre-pyramid Axis-wise Context Block, local-window Feature Guidance, assignment-aligned parent-level supervision, and training-time Poincaré path learning while preserving the detector’s native oriented-box formulation and leaf-level inference scoring. Under the common class-wise AP50 evaluation protocol, the proposed n- and s-scale variants achieved 77.5% and 82.7% on the filtered HRSC2016 test partition and 75.4% and 80.6% on the prepared ShipRSImageNet final-evaluation partition, respectively.