Semantic Segmentation With Global Encoding and Dilated Decoder in Street Scenes

Abstract

Semantic segmentation as a pixel-wise segmentation task provides rich object information, and it has been widely applied in many fields ranging from autonomous driving to medical image analysis. There are two main challenges on existing approaches: the first one is the obfuscation between objects resulted from the prediction of the network and the second one is the lack of localization accuracy. Hence, to tackle these challenges, we proposed global encoding module (GEModule) and dilated decoder module (DDModule). Specifically, the GEModule that integrated traditional dictionary learning and global semantic context information is to select discriminative features and improve performance. DDModule that combined dilated convolution and dense connection is used to decoder module and to refine the prediction results. We evaluated our proposed architecture on two public benchmarks, Cityscapes and CamVid data set. We conducted a series of ablation studies to exploit the effectiveness of each module, and our approach has achieved an intersection-over-union scores of 71.3% on the Cityscapes data set and 60.4% on the CamVid data set.