Probabilistic graphical modelling for semantic labelling of crowdsourced map data

Concerns regarding the accuracy of crowdsourced information limits its usage for several real world data-driven applications. In this paper we present a novel methodology for automated semantic prediction of street labels in crowdsourced maps. Toward the goal of finding best labels for streets, we use an undirected graphical model to capture three properties: the initial street labels given by the crowd as prior knowledge, the geometrical features of streets, and the inherent spatial relationships existing between streets in a network. Using the structural support vector machine paradigm a potential function is learnt on this model that jointly optimizes over the street labels in the entire network. We evaluate our methodology on the OpenStreetMap data for London and show that our model can predict 8 different street type labels with an accuracy of almost 90 percent. Our approach is more robust and improves upon the previous work where streets were assumed to have an independent and identical distribution.