News Video Story Segmentation with Multi-Modality Features

In this paper, we propose a detection-based method about the news video story segmentation. First, we extract the frames contained news title caption. Since every news story must have a counterpart title caption, and the news title caption always appears after the corresponding news story has begun for a few minutes. Then we can define a possible scope of story boundary candidates according to the time points of news title captions appearing. Finally, we select all the ends of the sentences in the possible scope as our boundary candidates. In this way, the story segmentation task can be converted into a boundary/non-boundary classification problem. To address this problem, we extract multimodal features, including lexical, acoustic, and visual features, and adopt a random forest trained with these features to accomplish the binary classification task. Several experiments are conducted to evaluate the performance of our proposed method, and the results suggest that the fusion of multi-modality features is beneficial to improve the performance of story boundary classification.