Eulerian Video Acceleration Magnification

The ability to amplify or reduce subtle image changesover time is useful in contexts such as video editing, medicalvideo analysis, product quality control and sports. In thesecontexts there is often large motion present which severelydistorts current video amplification methods that magnifychange linearly. In this work we propose a method to copewith large motions while still magnifying small changes. Wemake the following two observations: i) large motions arelinear on the temporal scale of the small changes; ii) smallchanges deviate from this linearity. We ignore linear motionand propose to magnify acceleration. Our method is pureEulerian and does not require any optical flow, temporalalignment or region annotations. We link temporal second-order derivative filtering to spatial acceleration magnifica-tion. We apply our method to moving objects where we showmotion magnification and color magnification. We providequantitative as well as qualitative evidence for our methodwhile comparing to the state-of-the-art.