As a hot research topic in the multimedia community, video analysis facilitates efficient video retrieval, browsing and management. It gives semantic tags to represent the relevant video content, which facilitate the following data management, retrieval and browsing. Previous methods usually analyze video content from a single modality, from which various low-level features are extracted to infer high-level semantics. Due to the existence of the semantic gap, such content-based methods can hardly extract detailed semantic descriptions from video content. Especially for the domain specific videos such as the sports video and the movie, the focus of the audience is not simple concepts (e.g. goal, quarrel), but instead, those detailed descriptions (such as Messi’s head goal, Ross is quarreling with Rach about the rent the living room). We propose cross-modality analysis to overcome the above difficulty. Specifically, we inspect the temporal correspondence between textual descriptions and video content, from which detailed semantics can be attached to the relevant video segment, and hence generating semantic video annotations. Based on the video analysis result, we propose two novel applications, personalized sports video customization and personalized movie scene synthesis,to meet audiences' personalized appetites. The former enables people to retrieve and summarize their interested video segments about specific player or event and the latter facilitate film producers to make their expected story movies through writting story scripts. Generally speaking, the main contributions of our work are as follows: 1. We propose a timestamp-independent method to annotate sports video content with external web text description. With the help of Bayesian network and keyword matching, video content and textual descriptions are first converted into semantic tag sequences. Each tag corresponds to a complete attack or an individual event. Then, sequence matching algorithm is used to align the above two semantic tag sequences and generate the final video annotations. 2. We realize a personalized sports video customization for mobile users. Considering the subjective content preferences and objective environment constraints, we raise a constraint optimization model to formulate the condition-limited video customization problem. Moreover, we design a concept social network to learn hidden user preference without adding additional user interaction. 3. We propose a novel ...
修改评论