Multimodal and Multilabel Genre Classification of Movie Trailers

Abstract

Movies are a diverse form of art and expressions. Unlike pictures and short clips, movies consist of a story-line which is deliberately made quite complex in order to engage the target audience. This paper presents evaluation of the usefulness of visual, textual and metadata-based functions for predicting the genre of a movie using movie trailers and analyzing it's visual features. The trailers were dissected into individual frames and were evaluated for key characteristics in order to divide them into different genres. Because previous articles employ an impractically large number of parameters to analyse the trailer, this approach aims to keep the number of parameters used to a minimum. The Moviescope dataset has been used which consists of about 5,000 movies with relevant information such as movie trailers, posters, plots and metadata. Linear Regression, KNN (K Nearest Neighbours), Decision Tree, Random Forest and Artificial Neural Networks are just a few of the classification algorithms this research study has used and compared.