EGBNet: a real-time edge-guided bilateral network for nighttime semantic segmentation

Due to poor illumination and low contrast, semantic segmentation of nighttime images faces major challenges. Various segmentation models with a large number of parameters are proposed to improve the performance but lead to an inability to process in real time. To tackle these problems, we propose a real-time edge-guided bilateral network (EGBNet) for nighttime semantic segmentation. Considering the blurred details and low contrast of nighttime images, we propose a lightweight multi-dilation dense aggregation module and introduce an efficient edge head to improve the ability to distinguish target features from the nighttime background. Moreover, a self-adaptive feature fusion module is proposed for the bilateral segmentation network to enhance the feature representation and generalization ability by fully using multi-scale feature maps. To capture more useful information from limited nighttime images, we further use the knowledge distillation strategy to improve the segmentation performance. Extensive experiments on ACDC and BDD datasets demonstrate the effectiveness of our EGBNet by achieving a satisfactory trade-off between segmentation accuracy and inference speed. Specifically, EGBNet achieves 55.56