Video completion based on improved belief propagation

In this paper, a novel computer vision technique is introduced to fill the holes in a video sequence due to removal moving foreground objects. Contrary to existing greedy techniques, we regard the task as a global optimization problem, which can overcome the obstacle of visual inconsistencies and don't require user intervention. After the pixel extensions in the frame via our definitions, we can deal with the time information included in video sequence freely and well. By means of minimizing the energy defined on the 3d label graph, we could match the appropriate patches to fill the hole. Since the traditional standard BP will lead to efficient problem caused by a great deal of labels, we improve the standard belief propagation algorithm and apply a better algorithm called priority belief propagation. This algorithm use label pruning and priority based message scheduling method to reduce the computational cost. From the experiment results, we can discover our method can considerably speed up BP's convergence and the completion performance is impressive.