Adaptive Reconstruction and Deformable Attention for UAV Small-Object Detection

Abstract

Detecting tiny objects in UAV imagery poses persistent challenges due to severe information loss during backbone downsampling, irregular target geometry, and the failure of global inference to resolve densely clustered micro-objects. We propose a three-component detection framework where each module addresses a specific failure mode and their interactions produce cumulative improvements across the pipeline. Cascaded Deformable Dual-Path Attention (CD-DPA) decouples edge-level structural cues from semantic channel responses to handle sub-32px irregular targets. A Confidence-Guided Adaptive SAHI pipeline identifies spatial detection failures via confidence-density scoring and selectively applies sliced inference only to weak regions, simultaneously feeding spatial failure masks to the reconstruction stream. Reconstruction-Guided Refinement (RGR) then converts backbone degradation into an actionable spatial prior that concentrates feature enhancement precisely on those flagged regions. Together, the full framework achieves 29.2 AP and 49.5 AP50 on VisDrone2019-DET, representing a 2.9-point gain over the Faster R-CNN baseline.