Shadow Removal for Enhanced Nighttime Driving Scene Generation

Abstract

Autonomous vehicles depend on robust vision systems capable of performing under diverse lighting conditions, yet existing models often exhibit substantial performance degradation when applied to nighttime scenarios after being trained exclusively on daytime data. This discrepancy arises from the lack of fine-grained details that characterize nighttime environments, such as shadows and varying light intensities. To address this gap, we introduce a targeted approach to shadow removal designed for driving scenes. By applying Partitioned Shadow Removal, an enhanced technique that refines shadow-affected areas, alongside image-to-image translation, we generate realistic nighttime scenes from daytime data. Experimental results indicate that our augmented nighttime scenes significantly enhance segmentation accuracy in shadow-impacted regions, thereby increasing model robustness under low-light conditions. Our findings highlight the value of Partitioned Shadow Removal as a practical data augmentation tool, adapted to address the unique challenges of applying shadow removal in driving scenes, thereby paving the way for improved nighttime performance in autonomous vehicle vision systems.