uICNet: Lightweight Image Segmentation

Abstract

Convolutional Neural Networks have started making headway in solving the problem of semantic segmentation. The demand for increasingly lightweight neural networks has been driven by an abundance of cheap hardware capable of running such neural networks and utilization of such networks for real world applications. MobileNet’ s utilization of the depthwise separable convolution has been proven to be an efficient approach for reducing neural network size without incurring a high penalty in accuracy. In the realm of image segmentation, ICNet was a breakthrough in the ability for semantic segmentation networks to be deployed on commonly available hardware and run at close to real time. In this paper, ICNet is improved upon by utilizing lessons learned from MobileNet and applying these lessons to create a new lighter weight network, uICNet. uICNet achieves similar accuracy to ICNet while substantially improving model size.