Semantic Labeling Enhanced by a Spatial Context Prior

Our observed visual world exhibits a structure, which implies that scene objects and their surroundings are not randomly arranged relative to each other but typically appear in a spatially correlated manner. Thus, the structural correlation can be exploited to make the visual recognition task predictable to a certain extent. Modeling relations between categories is, however, non-trivial, since categories are often represented at different granularities across distinct datasets. In this paper, we merge fine-level semantic descriptions into basic semantic classes which allows the generation of spatial contextual priors from a wide range of datasets. In this way, a contextual model is derived with the objective to employ the learned contextual prior to enhance visual recognition via improved semantic labeling. The prior is captured explicitly by computing occurrence and co-occurrence probabilities of specific semantic classes and class pairs from a diverse set of annotated datasets. We show improved semantic labeling accuracy by incorporating the contextual priors into the label inference process, which is evaluated and discussed on the Daimler Urban Segmentation 2014 dataset.