Movie Summarization via Sparse Graph Construction

We summarize full-length movies by creating shorter videos containing their most informative scenes. explore the hypothesis that a summary can be created assembling scenes which are turning points (TPs), i.e., key events in movie describe its storyline. propose model identifies TP building sparse graph represents relations between and is constructed using multimodal information. According to human judges, summaries our approach more complete, receive higher ratings, than outputs of sequence-based models general-purpose summarization algorithms. The induced graphs interpretable, displaying different topology for genres.