University of Sheffield at TRECVID 2008: Rushes Summarisation and Video Copy Detection

This paper presents our experiments on TRECVID workshop 2008. This year we participated in two challenging tasks, rushes video summarisation and video copy detection tasks. We studied a spatio-temporal video model to represent inter-frame correlation between two streams of video. The rushes summarisation task is aiming at creation of summaries with length not more than 2% of the original video. The approach involves the following three steps: irrelevant frame sequence removal, clapper shot detection, and shot alignment. The evaluation is made by human judges in four categories. It indicates that created summaries do not contain many duplications and junks. The summaries also have pleasant rhythm. The purpose of the video copy detection task is to detect pairs of copy (query) and reference video from the collection. We studied the problem with audio only queries that utilises the spatio-temporal video model.