Domain Adaptive Scene Classification Via Contrastive Learning And Uncertainty-Aware Fusion

Abstract

Scene classification is an important perception task for autonomous driving which allows for a higher level understanding of any traffic condition. Despite advancements in deep learning for scene classification, existing models struggle to generalize under domain shifts caused by geographic, lighting, and weather variations, leading to notable performance drops in real-world applications. With this challenge in mind, we propose a novel domain adaptive scene classification architecture to combine supervised contrastive learning with Bayesian uncertainty-aware feature fusion. The architecture first extracts global and local representations through an enhanced backbone model generated by Inception-V1 and Faster R-CNN to obtain class discriminative representations, however we are not limited in that we can use any backbone we choose. Next we also employ a momentum-encoded contrastive objective to align the representations and enhance the image representation space across source and target domains whereas there is no need to have any target labels. Finally, we developed a uncertainty-aware fusion module that uses Monte Carlo Dropout to weight the models predictions based on the confidence score from the model, as such, even if we the model has seen ambiguous or new domains we can use the collective models decision-making to maintain robustness for scene understanding from image representations. We provide comprehensive ablation studies on benchmark datasets (KITTI, BDD100K, Cityscapes) and real-world dashcam videos, demonstrating significant gains over state-of-the-art baselines in domain-level accuracy and generalization robustness. This work offers a strong basis for domain adaptive and reliable scene classification for use within safety critical applications in autonomous systems.