Density-Aware Adaptive Slicing Detection for Dense Small Objects in UAV Urban Low-Altitude Imagery

UAV aerial imagery for urban low-altitude sensing contains objects that occupy only a few pixels and are densely distributed, so whole-image inference at a limited input resolution represents extremely small objects inadequately. Fixed slicing raises recall but applies uniform parameters to every image, causing redundant computation in low-density regions and misses in high-density regions. We propose Density-Aware Adaptive Slicing Detection (DAASD): a low-threshold whole-image scan builds an object density heatmap, from which the slice size and overlap are selected, using small high-overlap slices for high-density regions and large low-overlap slices for low-density regions; per-slice detections are then mapped back and merged by non-maximum suppression. Using the end-to-end NMS-free YOLO26n on the full VisDrone validation set, DAASD raises the mAP50 from 31.34% to 37.51%, approaching the 37.66% of fixed slicing; on the first 500 test images, used as an additional cross-split reference, the mAP50 rises from 30.81% to 35.21%. It exceeds the 37.66% of fixed slicing. It also exceeds the retrained four-scale YOLO26n + P2 whole-image result of 32.82% by 4.69 points, indicating that preserving the local resolution at inference is more effective than adding a high-resolution feature layer. On the overhead xView dataset, where vehicles are as small as 12 pixels, slicing raises the mAP50 from 0.08% for whole-image inference to 69.19% without retuning the slice policy, confirming cross-dataset generalization. DAASD needs no retraining and has been validated on both the NMS-free YOLO26n and the RT-DETR detector, demonstrating its detector-agnostic nature.