Leveraging semantic saliency maps for query-specific video summarization

Automatic video summarization is a technique that allows us to easily understand and analyze large amount of visual data. Methods in the literature do this by dividing the video into certain parts and assigning a significance value to these parts. In query-based summary extraction, how query terms can be included in the importance value has been investigated. In this study, it was investigated which criteria could be effective in assigning importance value in query-based summarization, and it is suggested that semantic attributes and visual salience maps could contribute. In addition, these concepts are also included in the query terms, since the query terms may be the place where the activity is carried out or the actor performing the activity, rather than just some activities. In the experiments, it has been observed that the inclusion of semantic and visual saliency maps in the diversity term, which is one of the 3 functions included in the optimization, makes a positive contribution. Two other functions considered in optimization are representation and query term relevance. In short, with the thought that the places people look at the video may change when they view the video within a certain task (Task-Oriented Attention), it is useful i to predetermine the relevant parts in the video and to give extra weight to these sections in the summaries. Detailed experiments and visual results on two large datasets have demonstrated the correctness of the proposed hypothesis.