Uncertainty-Guided Query Emergence with Local Refinement Attention for Tiny Object Detection

2D tiny object detection is challenging because target objects often occupy only a few pixels and are easily overwhelmed by complex aerial backgrounds. While existing DETR-style detectors rely on fixed or density-based query allocation, such uniform allocation wastes decoder capacity on easy regions and under-invests in ambiguous ones. We propose replacing density-based control with prediction uncertainty as the active signal for query refinement. To address this limitation, we propose Uncertainty-Guided Query Emergence with Local Refinement Attention for vision-based tiny object detection. The proposed method strengthens fine-scale representation through multi-scale feature fusion and selectively refines ambiguous predictions using query-level uncertainty. Uncertainty is measured from classification entropy, localization instability, and confidence fluctuation. High-uncertainty parent queries generate bounded child refinement queries through Uncertainty-Guided Progressive Query Emergence. These child queries are then refined by Local Refinement Attention, which restricts attention to encoded memory tokens inside an enlarged support region around the parent box, reducing background interference while preserving transformer-level feature comparison. The complete UGQE-DETR framework is evaluated on VisDrone and AI-TOD-v2. On VisDrone, UGQE-S, UGQE-M, and UGQE-L achieve 35.8, 38.0, and 38.5 AP, respectively, with UGQE-M providing the best accuracy–cost balance. On AI-TOD-v2, UGQE-S, UGQE-M, and UGQE-L achieves 49.5, 57.7, and 64.9AP respectively. These results show that uncertainty-guided selective refinement can improve tiny object detection by directing computation toward unresolved hypotheses rather than processing all scene regions uniformly.