A video summarisation system for post-production
Post-production facilities deal with large amounts of digital video, which presents difficulties when tracking, managing and searching this material. Recent research work in image and video analysis promises to offer help in these tasks, but there is a gap between what these systems can provide and what users actually need. In particular the popular research models for indexing and retrieving visual data do not fit well with how users actually work. In this thesis we explore how image and video analysis can be applied to an online video collection to assist users in reviewing and searching for material faster, rather than purporting to do it for them. We introduce a framework for automatically generating static 2-dimen- sional storyboards from video sequences. The storyboard consists of a series of frames, one for each shot in the sequence, showing the principal objects and motions of the shot. The storyboards are rendered as vector images in a familiar comic book style, allowing them to be quickly viewed and understood. The process consists of three distinct steps: shot-change detection, object segmentation, and presentation. The nature of the video material encountered in a post-production fa- cility is quite different from other material such as television programmes. Video sequences such as commercials and music videos are highly dy- namic with very short shots, rapid transitions and ambiguous edits. Video is often heavily manipulated, causing difficulties for many video processing techniques. We study the performance of a variety of published shot-change de- tection algorithms on the type of highly dynamic video typically encoun- tered in post-production work. Finding their performance disappointing, we develop a novel algorithm for detecting cuts and fades that operates directly on Motion-JPEG compressed video, exploiting the DCT coeffi- cients to save computation. The algorithm shows superior performance on highly dynamic material while performing comparably to previous algorithms on other material.