ENet: A Deep Neural Network Architecture for Real-Time Semantic\n Segmentation

Abstract

The ability to perform pixel-wise semantic segmentation in real-time is of\nparamount importance in mobile applications. Recent deep neural networks aimed\nat this task have the disadvantage of requiring a large number of floating\npoint operations and have long run-times that hinder their usability. In this\npaper, we propose a novel deep neural network architecture named ENet\n(efficient neural network), created specifically for tasks requiring low\nlatency operation. ENet is up to 18$\\times$ faster, requires 75$\\times$ less\nFLOPs, has 79$\\times$ less parameters, and provides similar or better accuracy\nto existing models. We have tested it on CamVid, Cityscapes and SUN datasets\nand report on comparisons with existing state-of-the-art methods, and the\ntrade-offs between accuracy and processing time of a network. We present\nperformance measurements of the proposed architecture on embedded systems and\nsuggest possible software improvements that could make ENet even faster.\n