Video Trailer Generation using Multimodal Data Analysis

Trailers have been a vital part of the entertainment industry to pique audience interest. The process of trailer generation has been evolving to provide effective trailers to the public. In this paper we propose a novel approach by utilizing the audio and video modalities to extract key trailer segments that are compiled to create an effective trailer which highlights the impactful scenes of the short film. In contrast to previous work in this domain, the novel approach gives greater emphasis to the auditory features, since auditory features have a significant impact on the film in the horror genre. The dataset is a compilation of 311 short films and their respective trailers from various reputable public sources. Our approach introduces an audio guided visual model that compliments the acoustic features ensuring that the extracted segments are key moments both in terms of audio and video. The results obtained indicate that a significant proportion of predicted segments are deemed highly trailer-worthy.