Motivation. Semantic segmentation is a crucial computer vision task, solving which would enable a thorough scene understanding of the environment. The areas that already benefit from the automatic semantic segmentation include biomedical imaging [2], autonomous driving [4]. Further enhancement of current models will necessarily increase the number of possible applications, as well as quality of performance. The transfer learning of deep convolutional networks pre-trained for image classification on ImageNet has proven to be successful in semantic segmentation. In these models, last fullyconnected layers are replaced by convolutional ones followed by a learnable deconvolution or fixed interpolation to acquire the output of the same spatial size as the input. Usually, the segmented mask is coarse. Several ways to deal with this have been proposed, including the 'skip'-layer architecture [3] and post-processing with probabilistic graphical models [1]. Algorithm. The combination of graphical models with deep networks requires carefully designed differentiable operations to mimic approximate inference, while traditional upsampling approaches tend to operate only locally. To overcome these issues, we propose an alternative novel architecture aimed to perform an upsampling globally, as well as enforce the correct label recognition. For the first task, we propose the equivalent of deconvolution, which we call 'global interpolation'. We denote the decoded information of the RGB-image I : I ∈ R 3×H×W , as x : x ∈ R C×h×w , where C represents the number of channels, h and w define the reduced height H and width W , respectively. To acquire y : y ∈ R C×H×W , an upsampled signal, we apply the following formula:
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.