High-Resolution Semantic Labeling with Convolutional Neural Networks

摘要: Convolutional neural networks (CNNs) have received increasing attention over the last few years. They were initially conceived for image categorization, i.e., problem of assigning a semantic label to an entire input image. In this paper we address dense labeling, which consists in every pixel Since requires high spatial accuracy determine where labels are assigned, categorization CNNs, intended be highly robust local deformations, not directly applicable. By adapting networks, many labeling CNNs been recently proposed. Our first contribution is in-depth analysis these architectures. We establish desired properties ideal CNN, and assess how those methods stand with regard properties. observe that even though they provide competitive results, often underexploit could lead more effective efficient Out observations, then derive CNN framework specifically adapted problem. In addition learning features at different resolutions, it learns combine features. By integrating global information flexible manner, outperforms previous techniques. evaluate proposed compare state-of-the-art architectures on public benchmarks high-resolution aerial labeling.