Event-driven Automatic Editing for Videos on Desktop Manipulations - Effectiveness of Editing based on Behaviors of Drawing Attentions

摘要

In desktop manipulations such as cooking or assembly works, speakers usually draw, by typical behaviors, the audiences' attention to important portions. We verified the appropriateness and effectiveness of the video editing based on those behaviors in terms of statistics and comparisons with TV programs. In this paper, we first discuss significant types of human behaviors that commonly appear in presentations, and their usefulness as the triggers for camera (view) switching in TV programs. We show the advantages and drawbacks of this scheme by subjective evaluations. Then, we propose a simple and robust method for recognizing those behaviors, and demonstrate the efficacy by actual automatic editing on our video production system.