<p>Instance segmentation in autonomous driving faces significant challenges due to illumination variations, such as strong reflections, shadows, and occlusions. To address these issues, this paper introduces a novel polarization-aware framework that leverages pseudo-polarization information to enhance segmentation robustness under complex illumination conditions. Specifically, an unsupervised method is proposed to generate pseudo-polarization data, including degree of polarization and angle of polarization, directly from RGB images. These pseudo-polarization features are generated using optical physics constraints to ensure their physical plausibility and are further refined through multi-scale and multi-illumination consistency constraints, enhancing their reliability and semantic validity. The proposed polarization-aware network integrates these features into a model-agnostic architecture using an attention-based fusion mechanism, ensuring seamless compatibility with existing instance segmentation models. Extensive experiments demonstrate the effectiveness of the proposed method, achieving superior performance in scenarios with extreme illumination variations. Ablation studies validate the contributions of the pseudo-polarization generation process and consistency constraints, while real-world testing highlights its practical value for autonomous driving applications.</p>