Dual-inferences mechanism for real-time semantic segmentation

Autonomous cars have potential developments based on technology evolution. In the street scenes, the car needs to deal with a wide range of object sizes. Existing methods generally concentrate on deploying a single inference for semantic segmentation. However, one single scale is not suitable to capture the whole information of diverse sizes. It can effectively capture the context of thin objects, but it will get problems to cover the whole information of large objects, and reversely. In this paper, we propose an approach based on multi-scale inference to tackle the above difficulty. The multi-scale mechanism proposal employs two inference scales. Each scale is processed by a specific rate set of atrous spatial pyramid pooling. The segmentation maps are added together to take advantage of all scales. We validate our networks with a series of experiments on different open datasets. The approaches achieve high accuracy while reaching the speed for real-time semantic segmentation. The results are 75.5 % mIoU at 51 FPS on Cityscapes and 42.0 % mIoU on Mapillary Vistas.