movie2trailer: Unsupervised trailer generation using Anomaly detection

Abstract

In this work, we present movie2trailer - a novel unsupervised approach for automatic movie trailer generation. To our knowledge, it is the first-ever application of anomaly detection to such a creative and challenging part of the trailer creation process as a shot selection. One of the main advantages of our approach over the competitors is that it does not require any prior knowledge and extracts all needed information directly from the input movie. By leveraging the recent advancements in video and audio analysis, we produce high-quality movie trailers in equal or less time than professional movie editors. The proposed approach reaches state-of-the-art in terms of visual attractiveness and closeness to the “real” trailer. Moreover, it exposes new horizons for researching anomaly detection applications in the movie industry. The trailers, that were used in the evaluation stage are available at the following link - https:// bit.ly/ 2GbOj4R.

Tasks

Reproductions