Scan path and movie trailers for implicit annotation of videos

Abstract

Affective annotation of videos is important for video understanding, ranking, retrieval, and summarization. We present an approach that uses excerpts that appeared in the official trailers of movies, as training data. Total scan path is computed as a metric for emotional arousal, based on previous eye tracking research. Arousal level on trailer excerpts is modeled as a Gaussian distribution, and signed distance from the mean of this distribution is used to separate out exemplars of high and low emotional arousal in movies.