An improved belief propagation method for dynamic collage

This paper presents a new photo browsing technique, dynamic collage. Although previous photo collage techniques have innate advantages for viewing several photos collectively, they only focus on a static two-dimensional arrangement of photos so that the scalability is limited. In dynamic collage, new photos are incrementally inserted into the collage one by one while the old photos are removed accordingly, the positions of photos in the canvas are updated with a local and incremental manner to form a new layout so as to maximize visibility of all the important information embedded in the current collage meanwhile maintaining the visual continuity of two successive collages. To achieve this goal, a carefully designed optimization method based on belief propagation is employed. Unlike most traditional applications of belief propagation on pairwise MRF, we apply belief propagation on factor graph to optimize terms which cannot be represented by pairwise restricted belief propagation. We propose a novel approximate method to reduce the computation complexity, and this approximate method suggests a direction for using belief propagation on factor graph to optimize high order potential functions similar to ours.