Residual Conv-Deconv Grid Network for Semantic Segmentation

Abstract

This paper presents GridNet, a new Convolutional Neural Network (CNN)\narchitecture for semantic image segmentation (full scene labelling). Classical\nneural networks are implemented as one stream from the input to the output with\nsubsampling operators applied in the stream in order to reduce the feature maps\nsize and to increase the receptive field for the final prediction. However, for\nsemantic image segmentation, where the task consists in providing a semantic\nclass to each pixel of an image, feature maps reduction is harmful because it\nleads to a resolution loss in the output prediction. To tackle this problem,\nour GridNet follows a grid pattern allowing multiple interconnected streams to\nwork at different resolutions. We show that our network generalizes many well\nknown networks such as conv-deconv, residual or U-Net networks. GridNet is\ntrained from scratch and achieves competitive results on the Cityscapes\ndataset.\n