Four-level video summary coding

In this paper, we describe the video content using the codebooks of four summary levels, which are from lower to higher. The four levels are the local feature point, consistent region, object and scene. The local feature point is viewpoint and scale-invariant, and is described using local descriptor; the consistent region is the region that the color and texture remain consistent, it can be seen as a bag of local feature points; the object is composed by a number of consistent regions, and in this paper it is represented by the region adjacent matrix; the scene is the frame including some special objects and their space relationship, and it is represented by the object adjacent matrix. The codebook of the higher level is composed with the code from the codebook of the lower summary level. We report the experiment results in the ‘Groundhog Day’ movie.