A novel multi-metric scheme using dynamic time warping for similarity video clip search

In this paper, we describe an approach to video retrieval based on a multi-metric scheme. The video clip is represented by an ordered list of global frame features with systematic sampling. Similarity is measured by the combination order of feature vectors, which can be well described by the dynamic time warping method. However, it is still computationally expensive to make pairwise comparison in huge databases. To improve the search efficiency, we propose the category rate and dispersion rate as additional metrics of similar vector points to describe converge of origin series and filter out irrelevant candidates. A cheap-to-compute low bound estimate of dynamic time warping with Jaccard distance is also used to prune off unpromising candidates in KNN similar video search process. Experimental results on two benchmark databases show the efficiency of proposed approach.