Comparison of content selection methods for skimming rushes video

We compare two methods for selecting segments to be included in a video skim, using lists of relevant as well as redundant segments created from different visual features as input. One approach is rule-based, and creates a weighted sum of the input relevances. The other is HMM based, using a model trained on the TRECVID 2007 rushes data. The redundant segments are created from detection of repeated takes and junk content, the selected segments from visual activity and face detection. The results show that the approaches create very short summaries which only contain a part of the relevant information in the video, but reach very high scores in terms of the usability measures non-duplicates, non-junk and pleasant tempo. The HMM based approach contains more information despite shorter duration of the summaries.