ResSCNN: A semantic segmentation method for fast processing of large-scale input

Abstract

With the development of driver-assistance systems and driverless cars, vehicles are becoming more and more intelligent. However, today’s intelligent vehicles rely more on large sensors to sense the environment, so it is becoming more and more important for vehicles to be able to understand road information, and semantic segmentation has developed greatly. We find that the current mainstream semantic segmentation model is slow and inaccurate in the case of large scale input images. In this paper, we combine the large-scale input Fast-SCNN with the residual module to construct the residual semantic segmentation convolutional neural network (ResSCNN), after the experimental test, when the input scale is 1024 × 2048 px, yielding an accuracy of 68.4% mean intersection over union at 123.1 frames per second on Cityscapes dateset. Our network has achieved better results than the previous mainstream methods.