LoGF + GF, two kernels running in temporal and intensity domain, non-shifted response.
The shifted version in time domain equals to the phase change in frequency domain. The frequency of each sinusoidal component is shifted by the same amount as the time shift.
It is a subsequent work of the 2018 learning-based video motion magnification method. They use the same dataset. They generate the videos by manually overlapping two objects in the video and controlling the motion of different objects. Smaller implies raw video, larger implies magnified video. The goal is to learn the motion of the objects and magnify it. They focus on "motion/edge/phase", not "color/surface".
Short video segments with mild infant motion or those including caregiver shadow movement are classified as optimal. In contrast, videos with more significant movements, such as upper or lower limb movement (i.e., permissible motion artifacts), the presence of a parent or nurse’s hand that does not block the respiratory area, or mild crying that does not cause high motion artifacts, are considered suboptimal.
