DAFS-DETR: Dynamic-aware Adaptive Feature Selection Framework for Multi-scale Object Detection in UAV View

Small object detection from drone perspectives faces major challenges due to extremely small target sizes, large scale variations, and cluttered backgrounds. Existing detectors often suffer from limited shallow feature resolution, inefficient cross-scale semantic transfer, and inaccurate boundary localization, hindering precise and stable detection. This paper proposes DAFS-DETR (Dynamic-aware Adaptive Feature Selection for Multi-scale Object Detection), an enhanced RT-DETR framework integrating three key components: (1) a Dynamic Inception Feature Aggregation (DiFA) module for scale-adaptive feature extraction, reducing parameters while improving representation; (2) an Adaptive Sparse Attention-Enhanced Intrascale Feature Interaction Module (ASAE-IFIM) to suppress redundant background interference and enhance channel-wise information flow; and (3) a Recalibrated Boundary-Aware Feature Pyramid Network (RBA-FPN) for improved boundary feature aggregation and localization accuracy. Experiments on VisDrone2019 and UAVDT demonstrate that DAFS-DETR reduces parameters from 20.2M to 14.6M (approximately 27.7% reduction) while achieving mAP50 improvements of 2.9% and 1.0% and mAP50-95 improvements of 2.8% and 0.7% over RT-DETR-r18. These results validate that DAFS-DETR effectively addresses feature insufficiency, cross-scale transfer inefficiency, and boundary localization errors, achieving a superior balance of detection accuracy, robustness, and computational efficiency.