Differentiable Resolution Allocation for Tiny Object Detection in UAV Imagery

Aerial scenes exhibit a mismatch between where computation is spent and where information resides: targets occupy a small, unevenly distributed fraction of the frame, yet detectors process every region at identical resolution. Tiling and clustering pipelines exploit this sparsity, but the rule deciding which regions deserve magnification is hand-designed and never observes the detection loss. AFWD-Net removes that separation. A coarse density estimator, run on a heavily downsampled copy of the frame at negligible cost, produces a spatial prior over likely target locations; Gumbel top-K relaxation and a spatial-transformer sampler convert this prior into a small set of high-resolution crops whose centers and scales are differentiable, so the gradient that measures detection quality also determines where resolution is allocated. Within each crop, a learnable wavelet decomposition separates approximation from directional detail sub-bands and enhances the latter under edge supervision confined to annotated regions, restoring the high-frequency structure that stride downsampling erases, precisely the cue separating a genuine sub-32-pixel target from background clutter. Detection queries are seeded from the same density field rather than a fixed lattice, and crop-level predictions are merged by density-calibrated soft suppression. On VisDrone2019-DET the model attains 49.3% mAP@0.5, 30.4% mAP@[0.5:0.95] and 22.8% APS at 15.7 M parameters; a distilled 7.2 M student remains competitive with all compared methods. Experiments on UAVDT, AI-TOD-v2 and SODA-A assess cross-dataset generalization.