Automatic summarization of narrative video

The amount of digital video content available to users is rapidly increasing. Developments in computer, digital network, and storage technologies all contribute to broaden the offer of digital video. Only users’ attention and time remain scarce resources. Users face the problem of choosing the right content to watch among hundreds of potentially interesting offers. Video and audio have a dynamic nature: they cannot be properly perceived without considering their temporal dimension. This property makes it difficult to get a good idea of what a video item is about without watching it. Video previews aim at solving this issue by providing compact representations of video items that can help users making choices in massive content collections. This thesis is concerned with solving the problem of automatic creation of video previews. To allow fast and convenient content selection, a video preview should take into consideration more than thirty requirements that we have collected by analyzing related literature on video summarization and film production. The list has been completed with additional requirements elicited by interviewing end-users, experts and practitioners in the field of video editing and multimedia. This list represents our collection of user needs with respect to video previews. The requirements, presented from the point of view of the end-users, can be divided into seven categories: duration, continuity, priority, uniqueness, exclusion, structural, and temporal order. Duration requirements deal with the durations of the preview and its subparts. Continuity requirements request video previews to be as continuous as possible. Priority requirements indicate which content should be included in the preview to convey as much information as possible in the shortest time. Uniqueness requirements aim at maximizing the efficiency of the preview by minimizing redundancy. Exclusion requirements indicate which content should not be included in the preview. Structural requirements are concerned with the structural properties of video, while temporal order requirements set the order of the sequences included in the preview. Based on these requirements, we have introduced a formal model of video summarization specialized for the generation of video previews. The basic idea is to translate the requirements into score functions. Each score function is defined to have a non-positive value if a requirement is not met, and to increase depending on the degree of fulfillment of the requirement. A global objective function is then defined that combines all the score functions and the problem of generating a preview is translated into the problem of finding the parts of the initial content that maximize the objective function. Our solution approach is based on two main steps: preparation and selection. In the preparation step, the raw audiovisual data is analyzed and segmented into basic elements that are suitable for being included in a preview. The segmentation of the raw data is based on a shot-cut detection algorithm. In the selection step various content analysis algorithms are used to perform scene segmentation, advertisements detection and to extract numerical descriptors of the content that, introduced in the objective function, allow to estimate the quality of a video preview. The core part of the selection step is the optimization step that consists in searching the set of segments that maximizes the objective function in the space of all possible previews. Instead of solving the optimization problem exactly, an approximate solution is found by means of a local search algorithm using simulated annealing. We have performed a numerical evaluation of the quality of the solutions generated by our algorithm with respect to previews generated randomly or by selecting segments uniformly in time. The results on thirty content items have shown that the local search approach outperforms the other methods. However, based on this evaluation, we cannot conclude that the degree of fulfillment of the requirements achieved by our method satisfies the end-user needs completely. To validate our approach and assess end-user satisfaction, we conducted a user evaluation study in which we compared six aspects of previews generated using our algorithm to human-made previews and to previews generated by subsampling. The results have shown that previews generated using our optimization-based approach are not as good as manually made previews, but have higher quality than previews created using subsample. The differences between the previews are statistically significant