Abstract: To address the issue of suboptimal segmentation performance caused by the inadequate adaptability of current methods to nighttime scenes, this paper proposes a multi-level adversarial mean teacher network based on domain adaptation. The proposed methods segmentation process operates in two stages: Firstly, a curriculum style transfer strategy selects dusk scenes as the target style and transforms both daytime and nighttime images into dusk-style images. This approach decomposes the complex style transfer task into two simpler tasks, facilitating input style alignment. Subsequently, the multi-level adversarial mean teacher network performs adversarial learning at both the feature level and prediction probability level, achieving domain adaptation between the source and target domains across multiple levels and enhancing the models generalization capability across different domains. Additionally, the network employs dynamic class-domain mixing to introduce an extra mixed sample, enabling the model to learn richer dynamic class features. Experimental results demonstrate that the methods model achieves mIoU of 46.5%, 37.9%, and 47.8% on Dark Zurich, ACDC, and Nighttime Driving datasets, respectively. These findings indicate that the proposed method effectively improves the  adaptability and enhances the segmentation accuracy for nighttime urban landscapes.

Key words: domain adaptation, nighttime semantic segmentation, style transfer, adversarial learning, domain mixup