Interest Region Based Motion Magnification

摘要

In this paper, we proposed a method known as interest region based motion magnification for amplification of invisible motions. This method enables one to magnify subtle motion in the video for specific objects of interest to the user. To achieve this task, we have used object extraction using kernel K-means approach, automatic scribble drawing using super pixels and Bezier curves, alpha matting, and Eulerian motion magnification. The proposed method is tested on previously used video sequences for motion magnification and our own new videos with large background motion. We show the effectiveness of the proposed method by comparing with Eulerian motion magnification technique. We have presented visual results and performed no-reference video quality assessment for original videos and motion magnified videos. We further discuss the future improvements for motion magnification applications.