pixel-wise prediction

How can we train an end-to-end model for semantic segmentation? What is the role of architecture and how can we improve to get high resolution prediction maps? How can we deal with resource constraint environments?