Contextual deconvolution network for semantic segmentation

Abstract

In this paper, we propose a Contextual Deconvolution Network (CDN) and focus on context association in decoder network. Specifically, in upsampling path, we introduce two types of contextual modules to model the interdependencies of features in channel and spatial dimensions respectively. The channel contextual module captures image-level semantic information by aggregating the feature maps across spatial dimensions, and clarifies global ambiguity of features. Meanwhile, the spatial contextual module obtains patch-level semantic context by learning a spatial weight map, and enhance the feature discrimination. We embed the two contextual modules into individual components of the decoder network, thus improving the representation power and gaining more precise segment results. Thorough evaluations are performed on four challenging datasets, i.e., PASCAL VOC 2012, ADE20K, PASCAL-Context and Cityscapes dataset. Our approach achieves competitive performance with state-of-the-art models on PASCAL VOC 2012,ADE20K and Cityscapes dataset, and new state-of-the-art performance on PASCAL-Context dataset.