RDAQ: Real-Time DETR Meets Refinement-Driven Adaptive Querying for Dense Aerial Imagery

Abstract

Due to extreme density variations and scarce visual features, tiny-object detection in low-altitude uncrewed aerial vehicle (UAV) images remains a challenging task. Although DETR-based detectors have shown promising potential, they rely on a fixed number of object queries, which leads to suboptimal performance in dense scenes. To address this issue, we propose RDAQ, a density-aware dynamic detection framework. First, we design hierarchical squeeze fusion (HSF), which explicitly activates tiny-object details through top-down guidance, thereby bridging the semantic-spatial gap. Second, we design a density-guided routing estimation (DGRE) module to decouple query selection from static settings. By estimating the global crowding degree, DGRE adaptively assigns the optimal number of queries for each image. In addition, we propose a density-driven dynamic Filter (DDDF), which generates sample-specific convolutional weights according to density priors, significantly enhancing feature discriminability under severe occlusion. Extensive experiments on the DOTA-v1.0 and AI-TODv2 benchmarks demonstrate that RDAQ achieves state-of-the-art performance, obtaining $AP_{50}$ scores of 71.9% and 54.7%, respectively, thus providing a new paradigm for low-altitude image detection. Our code is available at https://github.com/Leanfawn/RDAQ