Automated Classification of Film Scenes based on Film Grammar

Abstract

Whenever there are extensive data collections to handle, metadata can help to structure and simplify this task. On websites for Video on Demand services or for film recommendations, we often find tags to describe and to search for films. However, manual tagging of films is a time-consuming task with a strong demand for human resources. In this paper, we present an approach for automated classification and indexing of film scenes. Although the classification algorithm uses audiovisual information, we mainly focus on visual information. The conception of the visual features is based on film grammar. In order to cover multiple aspects of film grammar, we enrich the row of state-of-the-art visual features by several novel ones. The evaluation of the classification framework is performed on five categories Suspense, Interaction, Essential Features, Dynamic and Valence. Altogether, the best classification rate equals an accuracy of 83.60% for the classification of the amount of dynamics in film scenes.