SGDA: A Saliency-Guided Domain Adaptation Network for Nighttime Semantic Segmentation

Abstract

Nighttime semantic segmentation has attracted considerable attention due to its crucial status in the smart city. However, it is challenging to handle poor illumination and indiscernible information. To tackle these problems, a saliency-guided domain adaptation network, SGDA, is proposed via adapting daytime models to nighttime scenes. Firstly, a saliency guidance branch is attached to the segmentation network to enrich the spatial features and guide the model to better perceive detail information. Secondly, to embed the saliency guidance to the segmentation network, a pyramid attention architecture is designed to fuse the features from the two branches. Thirdly, an illumination adaptation module is constructed to close the intensity distributions via adversarial learning, with an elaborately designed loss function to improve the performance. Extensive experiments on Dark Zurich dataset and Nighttime Driving dataset validate the effectiveness of SGDA, and indicate that our method improves the accuracy on small object categories,