Answers are grounded in this paper and cite their sources.
Oriented small object detection in remote sensing imagery faces the fundamentalSemantic-Geometric Decoupling Dilemma (SGDD), where entangled feature representations cause semantic information to be overwhelmed by geometric background interference, parameter coupling in oriented bounding box regression, and severe sample imbalance. To resolve this core challenge, we propose a Semantic-Geometric Decoupling and Refinement (TL-SGDR) framework with three key innovations. First, our semantic-geometry decoupled feature pyramid network addressesfeature-level SGDDby adaptively enhancing discriminative small-object To resolve this core challenge, we propose a Semantic-Geometric Decoupling and Refinement (TL-SGDR) framework with three key innovations. representations while suppressing background interference through context-aware fusion. Second, the semantic-customized prediction head eliminatesparameter-level SGDDby constructing dedicated feature spaces for position, size, and orientation parameters. Third, a geometry-aware optimization strategy mitigatessample-level SGDDby dynamically selecting samples based on inherent size-precision relationships. Extensive experiments on four challenging benchmarks, including DOTA-v1.0, DOTA-v1.5, HRSC2016, and DIOR-R, demonstrate state-of-the-art performance, achieving significant improvements of 1.73% mAP on DOTA-v1.5. Our work establishes a new paradigm for small object detection by fundamentally resolving the SGDD in remote sensing imagery. The code will be made publicly available at TL-SGDR.
No takes yet. Share an insight, caveat, or question.
Ning et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: