Zolghadr, E., Furht, B.
Abstract:
In this paper, a new framework for scene understanding using multi-modal high-ordered context-model is introduced. Spatial and semantical interactions are considered as sources of context and then incorporated in the model using a single object-scene relevance measure that quantifies high-order object relations. This score is used to minimize semantical inconsistencies among objects in dense graph representation of the scene category during the object recognition process. The new context model is then incorporated in a Conditional Random Fields (CRF) framework to combine contextual cues with appearance descriptors in order to increase object class prediction accuracy. A novel context-based non-central hypergeometric unary potential is defined to maximize the semantical coherence in the scene. Further refinement is performed using context-based pairwise and high-order potentials which use alpha-expansion and graph-cut to find optimal configuration. Comparison between the purposed approach and state-of-art algorithms shows effectiveness of this approach in annotation and interpretation of scenes.