Enhancing RT-DETRv2 with Early-Stage Feature Augmentation and Counting-Guided Dynamic Queries for Small Object Detection in UAV Images

Abstract

Objects in Unmanned Aerial Vehicle (UAV) imagery are typically small and densely distributed, where repeated downsampling weakens spatial details and a fixed query budget fails to accommodate large variations in scene density. To address these issues, we propose an enhanced RT-DETRv2 framework that incorporates early-stage feature augmentation and countingguided dynamic queries. Specifically, a high-resolution stride-4 feature map is reintroduced as a formal detection level within both the hybrid encoder and Transformer decoder, forming a four-level P2-P5 feature pyramid to preserve fine-grained spatial cues. Building on this design, we incorporate the counting-guided mechanism from DQ-DETR into our P2-enhanced RT-DETRv2 framework to jointly guide feature enhancement and adaptive query allocation. Experiments on the VisDrone2019-DET dataset at 640 × 640 resolution show that our method achieves $\mathbf{3 2. 0 7 {\%}}$ AP and 23.45% AP ${ }_{S}$, outperforming the RT-DETRv2 baseline by +4.15% AP, +5.42% AP ${ }_{S}$, +4.76% AP ${ }_{75}$, and +5.03% AR@100, while also surpassing D-FINE, YOLOv8m, and YOLO26m. These results validate that progressive enhancement from highresolution representation to density-aware feature augmentation and adaptive query allocation consistently benefits small object detection in UAV scenes.