Methods of Enriching The Flow of Information in The Real-Time Semantic Segmentation Using Deep Neural Networks

Abstract

Semantic Segmentation is one of the visual tasks that gained the significant boost in performance in recent years due to the popularization of Convolutional Neural Networks (CNNs). In this paper, we addressed the problem of losing information while changing the size of input images during training neural models. Moreover, our method of downsampling and upsampling could be easily injected into current autoencoder models. We show that without any significant changes in a model architecture it is possible to noticeably improve IoU metric. On popular Cityscapes benchmark, our model is achieving almost 2.5% boost in the accuracy of segmentation in comparison to the widely known ERF model. Additionally, to the ability to real-time usages, we run our network on GPU comparable to NVIDIA Jetson Tx2, what let us implement our algorithm in autonomous vehicles.