Improve the speed of semantic segmentation models from multiple perspectives

Abstract

The semantic segmentation model needs to output a category for each pixel, which is an intensive prediction task, so it needs more time. Real-world scenarios, such as autopilot, require fast reasoning speeds. Previous work has been to speed up model reasoning in one way or another, so the speed increase is limited. In this paper, we combine multi-level methods to speed up reasoning in a network and obtain remarkable results.