In this chapter, we have proposed a class of techniques named ViSig which are based on summarizing a video sequence by extracting the frames closest to a set of randomly selected SV's. By comparing the ViSig frames between two video sequences, we have obtained an unbiased estimate of their VVS. In applying the ViSig method to a large database, we have shown that the size of a ViSig depends on the desired fidelity of the measurements and the logarithm of the database size. In order to reconcile the difference between VVS and IVS, the SV's used must resemble the frame statistics of video in the target database. In addition, ViSig frames whose SV's are inside the VG should be avoided when comparing two ViSig's. We have proposed a ranking method to identify those SV's that are least likely to be inside the VG. By experimenting with a set of MPEG-7 test sequences and their artificially generated similar versions, we have demonstrated that IVS can be better approximated by using (a) SV's based on real images than uniformly random generation, and (b) the ranking method than the basic method. We have further characterized the retrieval performance of different ViSig methods based on a ground-truth set from a large set of web video.
The basic premise of our work is on the importance of IVS as a similarity measurement. IVS defines a general similarity measurement between two sets of objects endowed with a metric function. By using the ViSig method, we have demonstrated one particular application of IVS, which is to identify highly similar video sequences found on the World Wide Web. As such, we are currently investigating the use of IVS on other types of pattern matching and retrieval problems. We have also considered other aspects of the ViSig method in conjunction with its use on large databases. In our recent work , we have proposed a novel dimension reduction technique on signature data for fast similarity search, and a clustering algorithm on a database of signatures for improving retrieval performance.