Answers are grounded in this paper and cite their sources.
Power-line work-at-height monitoring requires the simultaneous detection of visually heterogeneous evidence: supervisory markers that may occupy only a few pixels after resizing, and worker-state cues that depend on body posture, equipment, and surrounding scene geometry. In a compact single-stage detector, this scale gap affects feature preservation, training assignment, and prediction stability. Shallow downsampling can weaken the high-frequency traces needed by tiny targets, standard matching may allocate too few positives to small categories, and different detection heads may produce inconsistent predictions for the same physical instance. We view these effects through the lens of asymmetric treatment at three stages of the detector, and introduce three targeted modifications to YOLO11n: a shallow wavelet detail preservation module that enhances low- and high-frequency sub-bands before resolution is lost; a class- and head-aware TinyAssign strategy that adjusts positive-sample allocation by category scale; and a ground-truth-aligned multi-scale consistency regularizer (GT-MSCR) that anchors cross-head agreement to ground-truth indices during training without inference overhead. On a self-collected four-class power-line dataset, and averaged over five independent runs, the model raises mAP@0.5 from 68.10% to 69.45% and recall from 60.84% to 65.32%. The largest per-class gain is obtained by the category most prone to scale-induced detection failure. External validation on the Pictor-v3 PPE and SH17 benchmarks further shows consistent gains on public data, with a parameter increase of only 0.005 M and an additional 2.66 ms of latency over the baseline.
No takes yet. Share an insight, caveat, or question.
Han et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: