Class Activation Map Stability as an Uncertain Guide to Denoising Gain for Frozen Pedestrian Detectors Under Gaussian Noise

Abstract Denoising can improve image quality without helping frozen pedestrian detectors. We ask whether the Attention-Preservation Score (APS), which correlates principal-component class activation maps of clean and processed images, tracks detection gain from denoising. We evaluate eight detectors, six pipelines and five Gaussian noise levels on 160 images, with geometry, polarity and attribution controls mainly on a reference detector, bootstrap intervals and an exploratory attention-guided pixel swap. Repairing incomplete patch coverage raises learned-denoiser APS by 0.09, detection change uncertain, and Block-Matching and 3D Filtering keeps the highest APS at the strongest noise in seven of eight detectors. APS is positively associated with detection gain for the reference detector, but pooled over detectors before repair the association is +0.02 (95% interval −0.15 to +0.17) and protocol-dependent; the primary swap interval includes zero. APS is a detector-specific stability measure, not a validated denoiser selector.