VIDEO STORY SEGMENTATION IN TRECVID 2004 Winston

In this technical report, we give an overview our technical developments in the story segmentation task in TRECVID 2004. Among them, we propose an information-theoretic framework, visual cue cluster construction (VC), to automatically discover adequate mid-level features. The problem is posed as mutual information maximization, through which optimal cue clusters are discovered to preserve the highest information about the semantic labels. We extend the Information Bottleneck framework to high-dimensional continuous features and further propose a projection method to map each video into probabilistic memberships over all the cue clusters. The biggest advantage of the proposed approach is to remove the dependence on the manual process in choosing the mid-level features and the huge labor cost involved in annotating the training corpus for training the detector of each mid-level feature. When tested in TRECVID 2004 news video story segmentation, the proposed approach achieves promising performance gain over representations derived from conventional clustering techniques and even the mid-level features selected manually; meanwhile, it achieved one of the top performances, F1=0.65, close to the highest performance, F1=0.69, by other groups. We also experiment with other promising visual features and continue investigating effective prosody features. The introduction of post-processing also provides practical improvements. Furthermore, the fusion from other modalities, such as speech prosody features and ASR-based segmentation scores are significant and have been confirmed again in this experiment.