Aspect Coherence for Graph-Based Semantic Image Labelling

Image semantic segmentation is the task of associating a semantic category label to each image pixel. This classification problem is characterised by pixel dependencies at different scales. On a small-scale pixel correlation is related to object instance sharing, while on a middleand large-scale to category co-presence and relative location constraints. The contribution of this paper is twofold. First, we present a semantic segmentation framework that jointly learns category appearances and smalland middle-scale pixel dependencies. The algorithm computational complexity is reduced by considering these two classes of dependencies separately. In particular, small-scale dependencies are accounted by clustering pixels into larger patches via image oversegmentation. To tackle middle-scale dependencies we propose a system based on a Conditional Random Field (CRF), that is, a discriminative graphical model, built over the patches. In this context, we propose a novel strategy to exploit local patch aspect coherence to impose an optimised structure in the graph that allows to have exact and efficient inference. The second contribution is a method