Seite - 27 - in Proceedings - OAGM & ARW Joint Workshop 2016 on "Computer Vision and Robotics“

Semantic Labeling Enhancedby aSpatialContext Prior DanielSteininger,Csaba Beleznai Austrian InstituteofTechnology,Austria Daniel.Steininger.fl@ait.ac.at Csaba.Beleznai@ait.ac.at Abstract Our observed visual world exhibits a structure, which implies that scene objects and their surround- ings 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 pre- dictable 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. Theprior iscapturedexplicitlybycomputingoccurrenceandco-occurrenceprob- abilitiesofspecificsemanticclassesandclasspairs fromadiversesetofannotateddatasets. Weshow improved semantic labeling accuracy by incorporating the contextual priors into the label inference process,which is evaluated anddiscussedon theDaimlerUrban Segmentation2014dataset. 1. Introduction Semantic segmentation of digital images links two core computer vision challenges: visual object recognitionandsegmentation. In recentyears, great improvement inaccuracies toboth task domains has been demonstrated, mainly due to a transition from learned hand-crafted representations towards representations distributed within hierarchies and embedded into compositional schemes, enabling a richgeneralization for a largenumber of object classes. Figure1. SemanticLabeling enhancedbya SpatialContextPriorand ConditionalRandom Field. 27