Clip based video summarization and ranking

In this paper, we present a new algorithm for video clip summarization and ranking, which is mainly based on a clip based video similarity measure and the affinity propagation clustering (AP) algorithm. We propose a proportional max-weighted bipartite matching algorithm for clip similarity measure. This method first generates a basic frame set and a corresponding proportion value set from each clip. Then it models two clips as a weighted bipartite graph, where the weight values are determined by both the direct frame similarities and the proportion values. Then the max-weighted bipartite matching is employed to measure the similarity between two clips. This method achieves good retrieval performance when the length of two clips varies greatly. With these clip similarities, clips are clustered using affinity propagation. The clips in one cluster generally describe the same video event. Video ranking is based on the cluster size and the average information entropy of each event. Experimental results are given to illustrate the proposed algorithm.