A Comparison of Score, Rank and Probability-Based Fusion Methods for Video Shot Retrieval

It is now accepted that the most effective video shot retrieval is based on indexing and retrieving clips using multiple, parallel modalities such as text-matching, image-matching and feature matching and then combining or fusing these parallel retrieval streams in some way. In this paper we investigate a range of fusion methods for combining based on multiple visual features (colour, edge and texture), for combining based on multiple visual examples in the query and for combining multiple modalities (text and visual). Using three TRECVid collections and the TRECVid search task, we specifically compare fusion methods based on normalised score and rank that use either the average, weighted average or maximum of retrieval results from a discrete Jelinek-Mercer smoothed language model. We also compare these results with a simple probability-based combination of the language model results that assumes all features and visual examples are fully independent.