Interactive Visual Object Extraction Based on Belief Propagation

Interactive foreground/background segmentation in a static image is a hot topic in image processing. Classical frameworks focus on providing one class label for the user to specify the foreground. This may be not enough in image editing. In this paper, we develop an interactive framework which can allow the user to label multiply foreground objects of interest. Our framework is constructed on belief propagation. The messages about the foreground objects and background are propagated between pixel grids. Finally, each pixel is assigned a class label after finishing the message propagation. Experimental results illustrate the validity of our method. In addition, some applications in color transfer, image completion and motion detection are given in this paper.