AO-SF: Alignment Optimization and Sample Fusion for Unsupervised Nighttime Segmentation

Among the existing nighttime domain adaptation methods, the semantic consistency of the same features is largely ignored in the process of semantic alignment, which leads to negative shift and distribution discrepancy between the source and target domains. To address this problem, we propose an unsupervised nighttime semantic segmentation network based on semantic alignment optimization and sample fusion (AO-SF), which can make a contribution to the segmentation performance of our nighttime model. Specially, the AO-SF constructs a nighttime generator with mutually exclusive classifiers in order to enhance the alignment of each feature. Moreover, we further design an image fusion model to leverage shared information between day-night image pairs, which can boost the quality of relighting network. The experimental results on Dark Zurich and Nighttime Driving show that the segmentation performance of our method is superior to that of other methods in unsupervised nighttime semantic segmentation tasks. The mIoU reaches 45.2% and 46.5%, respectively.