Film Shot Type Classification Based on Camera Movement Styles

Visual information contains the most important characteri-stics of a movie regarding the related content and filming techniques. Especially the way the camera moves to capture the scene is vital to define the director’s aesthetics. However, most of the machine learning tasks existing in the literature treat the movie as shallow content, rather than as an artistic work, and therefore focus on detecting objects and faces, recognizing activities and extracting plot-related topics. On the other hand, cinematography is closely connected to the choice of different ways to handle the camera, and thus camera movements include information that is useful in order to analyse the artistic style of a movie. In this work we present an original, publicly available (https://github.com/magcil/movie_shot_classification_dataset ) dataset for film shot type classification that is associated with the distinction across 10 types of camera movements that cover the vast majority of types of shots in real movies. In addition, two different methods are evaluated on the new dataset, one static that is based on feature statistics across frames, and one sequential that tries to predict the target class based on the input frame sequence using LSTMs. Based on the evaluation process it is inferred that the sequential method is more suited for modeling the camera movements.