Multi-Scenario Generalization via DVPCA and Reliability Analysis of Object Detection in Intelligent Systems

Abstract

Camera-based object detectors often suffer severe performance degradation under unseen adverse weather conditions, posing significant challenges to reliable autonomous driving perception. Existing physical-guided augmentation methods improve domain generalization by simulating realistic environmental degradations, yet their single-view optimization paradigm often induces feature-space domain separation and limits robustness across complex weather shifts. To address this problem, this paper proposes a Dual-View Physical Consistency Alignment framework for generalizable object detection, which constructs two symmetric physically degraded views for each training sample and enforces classification-level consistency while preserving localization independence, thereby encouraging weather-invariant semantic representation learning without sacrificing geometric precision. Experimental results on the Diverse Weather Dataset demonstrate that the proposed framework consistently outperforms existing state-of-the-art methods, improving the average unseen-domain mAP by 1.4 points over PhysAug. Notably, under the challenging Night Sunny domain, the proposed method achieves 47.1% mAP, outperforming the baseline by 13.6 points and PhysAug by 2.2 points. Furthermore, the mean performance under corruption is significantly increased from 33.3% to 41.3%. Visualization analysis further confirms enhanced weather-invariant feature alignment and more reliable object-centric attention activation.