MADA: Multi-Level Alignment in Domain Adaptation Network for Nighttime Semantic Segmentation

Semantic segmentation can provide spatial information about urban landscapes for autonomous driving. For safety consideration, segmentation methods should adapt to all day conditions including day and night. In this paper, a multilevel alignment in unsupervised domain adaptation network (MADA) is proposed for nighttime semantic segmentation. MADA aligns data distribution at multiple levels, utilizing the information of daytime source domain labels to improve the accuracy of nighttime semantic segmentation. Specifically, MADA includes two steps: 1) style-level alignment: twilight is chosen as intermediate domain and daytime and nighttime images are both transferred to twilight style; 2) feature-level and prediction-level alignment: discriminators are introduced after feature extractor and classifier in the segmentation network to align the distribution of features and predictions between the source and target domain images. The proposed MADA has achieved excellent nighttime segmentation results and minimized the training time. We conducted experiments on nighttime semantic segmentation datasets including Dark Zurich test set, ACDC nighttime validation set and Nighttime Driving test set. The mIoU reaches 44.3%, 37.59% and 44.92% respectively.