Answers are grounded in this paper and cite their sources.
Remote sensing object detection in low-altitude uncrewed aerial vehicle scenarios frequently involves densely distributed targets with arbitrary orientations and large viewpoint changes. Under such conditions, horizontal bounding boxes introduce severe background redundancy, ambiguous overlap, and inaccurate localization, making rotated object detection essential for precise geometry modeling and reliable instance separation. Meanwhile, frame-based sensors suffer from motion blur and limited dynamic range under high-speed maneuvering and drastic illumination variation, which weakens the contour continuity and directional cues required for accurate oriented bounding box regression. Event cameras provide asynchronous microsecond response and high dynamic range, but simply replacing the sensor is insufficient: after voxelization, event streams become sparse spatiotemporal tensors dominated by motion boundaries, while existing remote sensing benchmarks and rotated detectors are largely designed for dense appearance patterns. To address this gap, we construct the rotated event-based remote sensing object detection dataset (RERSOD), the first large-scale benchmark for this task in complex aerial observation conditions. Building on RERSOD, we propose the spatiotemporal-enhanced rotation detection network (STE-RDet), an event-aware rotated detector tailored to voxel-grid representations. STE-RDet explicitly addresses three key bottlenecks in event-based oriented detection: response dilution during upsampling, insufficient coupling between local structure and temporal dynamics, and background-dominated interference during multiscale aggregation. To this end, the framework constrains feature energy redistribution during upsampling, reinforces the interaction between local spatial topology and temporal event dynamics, and calibrates anisotropic context for orientation-sensitive regression. Extensive experiments on RERSOD demonstrate that STE-RDet achieves markedly better accuracy, robustness, and convergence efficiency than strong baselines and state-of-the-art rotated detectors.
No takes yet. Share an insight, caveat, or question.
Qiao et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: