Semantic segmentation requires both rich spatial information and sizeablereceptive field. However, modern approaches usually compromise spatialresolution to achieve real-time inference speed, which leads to poorperformance. In this paper, we address this dilemma with a novel BilateralSegmentation Network (BiSeNet). We first design a Spatial Path with a smallstride to preserve the spatial information and generate high-resolutionfeatures. Meanwhile, a Context Path with a fast downsampling strategy isemployed to obtain sufficient receptive field. On top of the two paths, weintroduce a new Feature Fusion Module to combine features efficiently. Theproposed architecture makes a right balance between the speed and segmentationperformance on Cityscapes, CamVid, and COCO-Stuff datasets. Specifically, for a2048x1024 input, we achieve 68.4% Mean IOU on the Cityscapes test dataset withspeed of 105 FPS on one NVIDIA Titan XP card, which is significantly fasterthan the existing methods with comparable performance.