A Computable Visual Attention Model for Video Skimming

A novel computable visual attention model (VAM) for video skimming algorithm is proposed. Videos bear more motion features than images do. Objects in videos cause different attention effects, depending on various situations, positions, motions, and appearances. The static visual attention model is based on spatial distribution, visual object, or both, but fall short in solving temporal attention effects. The proposed VAM model adopts the alive-time(AT) of a visual object as a new descriptor to improve the accuracy of locating highlight in a video clip, then produces better video skimming results. The model is represented by a set of descriptors to be computable and provide a generic framework for video analysis. The temporal variations of attention value in a video clip are weighted by non-linear Chi-square distribution. Then the highlights of the frames in the video are represented by the attention window (AW) and the attention values of the visual objects (AOs) are tracked and used to generate the attention curve of the video. At last, a video skimming strategy is used to select the highlights of the video by analyzing the attention curve. The experiment result shows that the proposed model makes the skimming results 15%~25% shorter than previous methods.