Visual-semantic collaborative interactions in synergic system for scene modeling

With the reference of slaving principles from synergetic in physics, the embedding correspondences from visual cues and semantic ones are captured as prototypes in dynamic system. In scene modeling for interpretation, collaborative visual-semantic interactions are factorized as self-stimulation, self-restraint and other-restraint processes with decrease of energy to stable situation. The experiments demonstrate our framework can automatically correct false local labels and improve the accuracy of scene categorization.