Abstract
This paper presents a novel match-and-tiling approach to retrieve video sequences. The approach considers video similarity matching at two levels -- the shot and sequence levels. At the shot level, we transform the matching of similar shots into a problem of matching video feature trajectories using a longest sub-sequence matching technique. This is to achieve both computational simplicity and retrieval effectiveness. At the sequence level, we view sequence matching as a clustering problem and employ an effective sliding window algorithm to locate multi-occurrences of similar video sequences in the database. The resulting technique is able to retrieve both exact and similar video sequences with different durations and shot ordering. Our results demonstrate that our technique is robust and effective in retrieving similar video sequences.