Answers are grounded in this paper and cite their sources.
Recent advances in Arbitrary-Oriented Object Detection (AOOD) for remote sensing imagery have largely relied on high-capacity architectures, such as two-stage detectors and transformer-based models, to handle arbitrary object orientations and extreme aspect ratios. Although these approaches achieve high detection accuracy, they often incur substantial computational overhead, limiting their applicability in real time scenarios. In contrast, one-stage detectors offer superior efficiency, but typically suffer from insufficient rotation-aware feature representation, resulting in suboptimal localization accuracy. We introduce Rotation-aware Deformable Convolution Network (RDCNet), an efficient one-stage anchor-free architecture for AOOD. RDCNet incorporates a rotation-aware deformable convolution module into the backbone, where offset sampling is formulated in polar coordinates to decouple scale and orientation, enabling stable encoding of rotational information at early-stage feature maps. This rotation-aware representation is propagated through deeper layers enriched with semantic information. Furthermore, we introduce an Aspect-Adaptive Label Assignment (AALA) strategy that integrates an aspect-agnostic centerness term to guide positive sample selection, effectively reducing learning bias toward objects with extreme aspect ratios. Extensive experiments on AOOD benchmarks, including DOTA-v1.0, DOTA-v1.5, and HRSC2016, demonstrate that RDCNet achieves state-of-the-art or highly competitive performance among one stage detectors while maintaining real-time inference speed, and exhibits favorable efficiency–accuracy trade-offs compared to representative two-stage approaches. These results indicate that RDCNet provides a practical and effective solution for AOOD in complex remote sensing scenes.
No takes yet. Share an insight, caveat, or question.
Ryu et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: