SAF-RTDETR: Spatially Aligned High-Resolution Feature Fusion for Small Object Detection in UAV Imagery

Abstract Backbone downsampling weakens small objects in unmanned aerial vehicle imagery, whereas direct use of shallow features also amplifies structured background. We propose SAF-RTDETR, a four-scale RT-DETR extension that treats high-resolution detection as selective feature fusion. A grouped sampling unit aligns stride-8 shallow and semantic features; the aligned feature then guides a directionally weighted stride-4 branch. Partial-channel single-head attention supplies high-level context. Under the same 100-epoch VisDrone2019-DET protocol, three runs obtain 31.14 ± 0.51% AP50:95 versus 28.46 ± 0.73% for RT-DETR-R50, with gains in every paired seed and a 2.77-point increase in APS. The model also leads all 18 tested corruption conditions and raises mean corrupted AP from 24.85% to 27.45%. Cross-dataset AP changes are +4.54, +1.00, and −0.34 points on AI-TOD-v2, HIT-UAV, and UAVDT. SAF-RTDETR reaches 57.62 FPS on an RTX 5090 but requires 288.23 GFLOPs. The results support selective high-resolution fusion while exposing its computational and transfer limits. A controlled ablation further shows that a plain stride-4 level adds only 0.63 AP, whereas selective enhancement accounts for most of the gain.