Semantic segmentation requires both large region context information and rich spatial information. The former can be achieved by a deep-enough network structure. However, most spatial information lost in repeated down-sample operation in deep Convolutional Neural Networks cannot be easily recovered. In this paper, we propose a novel edge-guided hierarchically nested network to improve both speed and accuracy of real-time semantic segmentation. We introduce Edge Guidance Unit which aims to guide the decoder network to exploit high-resolution clues, recover and refine spatial information with low additional computational cost. We also introduce a well-designed Fusion Block to combine different resolution features along with edge guidance. In experiments we demonstrate that our network stabilizes the small object region while achieving a result of 72.7% mean IoU at 40 FPS on CityScapes test dataset.