Smooth and clear : auxiliary domain and inter-class contrast adaptation for semantic foggy scene segmentation

Abstract

Semantic segmentation, a crucial task in computer vision, classifies pixels in an image into predefined set of objects for a diverse range of applications including autonomous driving. Adverse conditions, in particular due to fog, pose significant challenges to the models trained on clean data for autonomous driving. Unsupervised Domain Adaptation (UDA) methods aimed to address these challenges have struggled with large domain gaps caused by fog-induced shifts. Additionally, fog makes it difficult for a model to distinguish boundaries between classes clearly. Toward overcoming these challenges, our study introduces a novel framework utilizing auxiliary domains — foggy source and clean target. We leverage the foggy source to adapt the model to foggy conditions smoothly while we exploit the clean target to impart knowledge about diverse data styles. Additionally, we address class boundary blurring in foggy images through inter-class contrast adaptation (ICCA). Our approach achieves state-of-the-art performance by effectively utilizing auxiliary domains and ICCA in standard settings even with a lightweight baseline.