RBDYOLO improves small object detection in low altitude remote sensing imagery
Lulu Liu, Rui Zhu, 郭辰嘉, Yushuai Zhang, Zhibo Shi, Yaru Li, Le Gao · Scientific Reports · 2026
Unmanned aerial vehicle (UAV) imagery at low altitude is increasingly used for traffic oversight, infrastructure inspection, and emergency awareness, yet bird's-eye views exhibit pronounced scale diversity, minute object footprints, and cluttered backgrounds. Default YOLO-style single-stage detectors on a P3-P5 pyramid can therefore suffer coupled cross-scale misalignment and within-scale receptive-field mismatch, increasing misses and unstable calibration on tiny instances. Built on the Ultralytics YOLO11 macro graph and training recipe for strict comparability, RBD-YOLO addresses these bottlenecks with three coordinated methods: (R) within-scale adaptive receptive-field reweighting, (B) learnable cross-scale weighted fusion, and (D) scale-space-task joint dynamic calibration. Under one unified protocol, with VisDrone-DET validation as the primary benchmark and UAVDT as an auxiliary cross-dataset check, RBD-YOLO achieves 49.6% precision, 39.4% recall, and 40.1% [email protected] on VisDrone, outperforming the same-series YOLOv11n baseline (44.4%, 33.2%, 33.1%) and remaining ahead of newly added larger-scale YOLO-family comparisons in [email protected]. On UAVDT, it reaches 98.2% [email protected] and 95.4% recall, indicating competitive cross-dataset behavior on a nearly saturated vehicle-centric benchmark. Ablations attribute decomposable gains to methods B and R, a dominant increment from widening fusion bandwidth, and further refinement from method D once cross-scale capacity is sufficient; at 640 × 640 the full model uses about 3.47 M parameters and 16.4 GFLOPs, yielding an interpretable accuracy-complexity trade-off for VisDrone-like aerial small-object detection.