GC-DETR: Geometry Conditioned Real-Time Object Detection With Enhanced Quality Consistency for Open Water in UAV Imagery
Abstract
Real-time and reliable object detection is essential for uncrewed aerial vehicle (UAV)-based open-water search and rescue, where delayed or missed detections may directly affect situational awareness and rescue decision-making. However, variations in UAV altitude, attitude, and viewing perspective continuously alter target scale and spatial distribution, while water-surface reflections, complex background textures, and weak small-object features further degrade detection reliability. In detection transformer (DETR)-based detectors, classification scores do not directly reflect bounding-box localization quality, allowing poorly localized predictions to receive high confidence and reducing the reliability of prediction ranking. To address these challenges, we propose GC-DETR, a geometry-conditioned real-time detector for UAV-based open-water scenarios. The sensor-aware query module (SAQM) encodes UAV flight-state information to condition object queries, enabling query representations to adapt to changing imaging geometries with limited computational overhead. The spectral decomposition adaptive decoupled attention (SDADA) module combines spectral decomposition, adaptive sampling, and head-wise geometric decoupling to improve cross-scale feature alignment, preserve weak target cues, and suppress complex water-surface interference. The IoU perception quality head (IOUPQH) explicitly estimates localization quality and calibrates classification confidence, thereby improving the consistency between prediction scores and bounding-box accuracy. Experiments on SeaDronesSee and additional aerial datasets demonstrate that GC-DETR improves $\mathrm{AP}_{50:95}$ and APSmall on SeaDronesSee by 1.85 and 3.57 percentage points, respectively, while reducing the parameter count by 16.3%.attains a 2.7% $\mathrm{AP}_{50:95}$ gain on a terrestrial aerial traffic dataset. These results show that GC-DETR achieves a favorable balance between detection accuracy and model efficiency, providing a more reliable solution for UAV-based open-water search-and-rescue (SAR) operations under changing flight conditions. The code will be available at https://github.com/shuaimadhu/GC-DETR