SCW-DETR: UAV small object detection with cross-spatial-column calibration and structurally-constrained representation

Abstract Missed detections of small objects in object detection are fundamentally similar to the selective-ignorance phenomenon in human vision. One possible explanation is that small-object evidence is gradually weakened across the detection pipeline: weak target responses may be suppressed by structured background textures during feature extraction; shallow boundary cues and deep semantic cues may become misaligned during cross-scale fusion; and small pixel deviations can have a larger relative impact during bounding-box regression. To address this error chain, we propose SCW-DETR and design three collaborative components around the process of evidence generation, transmission, and anchoring. First, a cross-spatial-column calibration block is introduced to enhance weak target responses in cluttered backgrounds by combining column-wise calibration, spatial attention, and self-calibration, while a feature-splitting strategy is used to control parameter growth. Second, a structurally constrained representation Network is constructed to reduce unreliable responses and improve the interaction between semantic and boundary information during multi-scale propagation. Finally, a wise-focaler-shapeIoU loss is designed to combine dynamic weighting, hard-sample focusing, and shape constraints for more stable bounding-box regression. Experimental results on VisDrone2019, AI-TOD, HIT-unmanned aerial vehicle (UAV), and RSOD demonstrate the effectiveness of the proposed method. On VisDrone2019, SCW-DETR achieves an mAP50 of 53.4%, which is 6.0% points higher than the baseline. On AI-TOD, HIT-UAV, and RSOD, SCW-DETR reaches mAP50 values of 52.8%, 82.2%, and 94.4%, improving the baseline by 2.3, 3.9, and 3.7 percentage points, respectively.