A content-based approach for detecting highlights in action movies

Although detecting highlights in films is a trivial task for humans, previous studies have not determined whether a computer can be equipped with this capability. In this paper, we present a content-based system that automatically detects highlight scenes and predicts highlight scores in action movies. In particular, high-level image attributes and an early event detection approach are applied. Dissimilar to current learning-based approaches that model the relationship between the whole highlight and corresponding audiovisual features, the proposed system studies the temporal changes of a set of general features from a nonhighlight to a highlight scene. The experimental results indicate that achieving the highlight detection task is technically feasible. It also provides critical insights into understanding the feasibility of solving this challenging problem. For example, both audio and visual features are crucial and the filming style can be captured using high-level image attributes, which further improve the overall detection performance.