Overview of the semantic segmentation network. The first part of our network calculates a feature for each input pixel by exploiting a fully convolutional network (FCN) with in-network upsampling and skip-and-fuse architecture to fuse coarse, semantic, and local, appearance information. The second part of the network adds binary potentials (i.e., adding constraints to give neighboring pixels with similar intensity the same label) by using the dense CRF-RNN as proposed by