Highlighting action content in animated movies

In this paper we tackle the issue of highlighting action content in animated movies, which proves to be a valuable information for retrieving movies in content-based video indexing systems. We use the hypothesis that action is in general related to a high frequency of video transitions, and adapt it to the constraints of the animation domain. First, we perform a video temporal segmentation by detecting cuts, fades, dissolves and specific color effects. Second, we analyze the movie rhythm, in terms of shot changes over a time unit. We target several action categories, namely: hot action, regular action and low action. This constitutes the action groundtruth. Finally, we employ a four step algorithm (thresholding, merging, pruning, restoring complementarity) to highlight movie parts according to the previously determined groundtruth. The efficiency of our approach was tested on several pre-labeled animated movies, achieving precision and recall ratios of more than 70%.