АДАПТИВНИЙ КАСКАД ЗА СКЛАДНІСТЮ СЦЕНИ ДЛЯ ВИЯВЛЕННЯ ОБ’ЄКТІВ НА АЕРОЗНІМКАХ БЕЗПІЛОТНИКІВ: КОЛИ ВИБІРКОВА ЕСКАЛАЦІЯ ОБЧИСЛЕНЬ ВИПРАВДАНА

Abstract

Unmanned aerial vehicles are a standard video-monitoring instrument in security tasks — from perimeter and critical-infrastructure protection to situational awareness — and detector ensembles raise the accuracy of object detection in aerial imagery at a manifold per-frame compute cost, which is critical for onboard systems. A Scene-Complexity Adaptive Cascade (SCAC) is proposed and investigated: a cheap first-tier detector processes every frame and computes a dimensionless, annotation-free complexity signal; only frames whose complexity exceeds a threshold are escalated to an expensive second-tier WBF ensemble. The threshold is not hand-tuned — it is calibrated by cross-fitting at the video-sequence level for a target average budget. The methodological contribution is an oracle feasibility gate: a greedy reference curve ordered by the realized per-frame ensemble gain is constructed from cached predictions, and realizable complexity signals are compared against it. In the training domain (VisDrone-DET) the ensemble gain is dense across frames (75–90 % of frames improve), so realizable complexity signals do not beat uniform fusion at the same budget, and the popular uncertainty signal is consistently worse than random escalation. Under domain shift (zero-shot transfer to UAVDT with official condition labels), by contrast, the ensemble gain is sparse (41 % of frames; at night the ensemble hurts), and the cross-fitted cascade with a detection-density signal reaches the quality of the full ensemble (0.618 mAP@50, scene-balanced) at 45 % of its compute, significantly outperforming the single model (+0.022 on the full sample; bootstrap mean +0.017, 95 % CI [+0.002; +0.032]) and random escalation, including its scene-stratified variant (+0.019 on the full sample; bootstrap mean +0.016, 95 % CI [+0.001; +0.029]). The realized escalation profile (9 % night / 22 % day / 45 % fog) reproduces scene adaptivity without any scene classifier, bypassing the known bottleneck of routing — the unrecognizability of adverse conditions from frame appearance. A practical criterion is formulated: selective compute escalation pays off only where the expensive model's gain is sparse across frames — typically under domain shift, not in the training domain.