Cognitive Compensation Object Detection for Single-Source Domain Generalization in Complex Weather

Abstract

Complex weather changes road-image visibility, spectral statistics, and local structure, making detectors trained on clear daytime data unreliable across unseen fog, rain, night, and night-rain domains and limiting autonomous driving, roadside perception, and intelligent-transportation vision. Existing object-centric dynamic detectors use visionlanguage prompts to select target-relevant features but do not explicitly model fine-grained weather degradation or the complementary roles of training-stage weather experience and test-stage environmental compensation. Inspired by human vision integrating accumulated experience with scenedependent compensation, we propose Cognitive Compensation Object Detection (CCOD) for single-source domaingeneralized object detection in complex weather. CCOD includes an Experience Generalization Module (EGM) that builds diverse weather experiences via frequency-domain perturbation and physics-based degradation while enforcing weather-invariant object representations, and an Environmental Compensation Module (ECM) that constructs reference views for unlabeled test images and filters reliable detections via multi-view consensus. On Diverse-Weather, CCOD increases average target-domain AP50 to 35.82%, outperforming PDOC by 3.52 points and validating joint experience generalization and environmental compensation.