Video-similarity search based on mixture distributions

A method is described for searching for videos similar to a query. The video frames are segmented into shots by clustering consecutive frames based on their color histograms. Segmentation performance is improved by adding temporal information to the color data. Videos are represented by their temporal distribution of the colors, and the distance between videos is estimated based on the effort needed to transform one distribution into another. The distance is used to search videos similar to a query. The search is speeded up by using a filtering method based on the inequalities in distances. Testing of the method showed that the performance of the similarity search is improved by incorporating the temporal information.