Three-branch Semantic Segmentation Network

Abstract

Effective Semantic segmentation requires both spatial details and object-level semantic information. Meanwhile, context information is also important for complex scene understanding. However, it is hard to meet these demands simultaneously in the top-down CNN structure. In this paper, we tackle this problem with a Three-branch Semantic Segmentation Network (TSS Net). The proposed TSS Net consists of three parts, including a spatial network, a semantic network and a context network. The spatial network utilizes convolutional layers with small stride and a spatial pyramid pooling module to extract multi-scale spatial features. In the semantic network, multiple level features are combined to enhance semantic information. The context network integrates different scales contextual information to facilitate objects localization in complex scene. The proposed semantic segmentation framework has been evaluated on the CamVid and the Cityscapes datasets. Experimental results demonstrate that the TSS Net achieves state-of-the-art performance.