Abstract
To help users select movies they want to watch, we have developed a new algorithm for semi-automatically generating movie summaries. The algorithm is based on an empirical study of how existing movies and trailers are created. Applying the Semantic Score method to capture the movie's development quantitatively, we have identified several constraints that can be used to build movie trailers. These include : 1) Select more scenes from the lntroduction parts than the Development and Turn parts of the movie, and no scenes from the Conclusion. 2) For scenes in the lntroduction part of the movie, select those scenes that have an absolute Semantic Score higher than a threshold, and those that occur at local maxima (peaks) in the Semantic Graph, and those that follow local maxima. 3) For scenes in the Development and the Turn parts of the movie, select those scenes whose score is higher than a given threshold and those that occur at local maxima. 4) Otherwise, choose pictorial scenes of low score as needed. Furthermore, the signal processing based shot detection and cutting are also taken into consideration. We have evaluated this approach experimentally by showing semi-automatically generated summaries to viewers. The results showed that first the users rated the movie summaries favorably, both as summary and as creative work in its own right, and second that the generated movie summaries enhanced the viewer's desire to watch the original movie.
