Efficient Decoder and Intermediate Domain for Semantic Segmentation in Adverse Conditions

Abstract

In smart city contexts, traditional methods for semantic segmentation are affected by adverse conditions, such as rain, fog, or darkness. One challenge is the limited availability of semantic segmentation datasets, specifically for autonomous driving in adverse conditions, and the high cost of labeling such datasets. To address this problem, unsupervised domain adaptation (UDA) is commonly employed. In UDA, the source domain contains data from good weather conditions, while the target domain contains data from adverse weather conditions. The Adverse Conditions Dataset with Correspondences (ACDC) provides reference images taken at different times but in the same location, which can serve as an intermediate domain, offering additional semantic information. In this study, we introduce a method that leverages both the intermediate domain and frequency information to improve semantic segmentation in smart city environments. Specifically, we extract the region with the largest difference in standard deviation and entropy values from the reference image as the intermediate domain. Secondly, we introduce the Fourier Exponential Decreasing Sampling (FEDS) algorithm to facilitate more reasonable learning of frequency domain information. Finally, we design an efficient decoder network that outperforms the DAFormer network by reducing network parameters by 28.00%. When compared to the DAFormer work, our proposed approach demonstrates significant performance improvements, increasing by 6.77%, 5.34%, 6.36%, and 5.93% in mean Intersection over Union (mIoU) for Cityscapes to ACDC night, foggy, rainy, and snowy, respectively.