Semantic segmentation of urban scenes by learning local class interactions

摘要: Traditionally, land-cover mapping from remote sensing images is performed by classifying each atomic region in the image isolation and enforcing simple smoothing priors via random fields models as two independent steps. In this paper, we propose to model segmentation problem a discriminatively trained Conditional Random Field (CRF). To end, employ Structured Support Vector Machines (SSVM) learn weights of an informative set appearance descriptors jointly with local class interactions. We principled strategy pairwise potentials encoding preferences sparsely annotated ground truth. show that approach outperform standard baselines more expressive CRF models, improving 4–6 points average accuracy on challenging dataset involving urban high resolution satellite imagery.