Multi-path Fusion Network For Semantic Image Segmentation

Abstract

Recently, deep convolutional neural networks (CNNs) have led to significant improvement over semantic image segmentation and have also been the best choice. In this paper, we propose a deep neural network architecture, Multi-Path Fusion Network (MPFNet), for semantic image segmentation. In MPFNet, we add more convolution paths to every convolution layer. The depth of each convolutional path increases linearly, which provides a superior method for pixel level prediction. Using this method, we integrate contextual information and local information to produce good quality results on the semantic segmentation task. In addition, dense skip connections are added to repeatedly leverage previous features. The proposed approach improves strong baselines built upon VGG16 on two urban scene datasets, CamVid and Cityscapes, which demonstrate its effectiveness in modeling context information.