Multi Domain Learning for Motion Magnification

Figure 1. Hand Drill: Magnifying rotational motion is a difficult task. So to evaluate SOTA methods (b) Acceleration method [29], (c) Jerk-aware [24], (d) Anisotropy [22], (e) Oh et al. [17], and the proposed method (f) D1, (g) D2, a video containing a hand drill with rotational motion along its axis is used. In 2D, this motion is visible as a spiral motion. So, magnification can be perceived as an increase in spiral motion (shown in spatial-temporal slices taken from the red strip at the right part of the figure). Hand-crafted methods [22], [24], [29] have small magnification (less outward radius in temporal slices) and produce ringing artifacts (visible as white edges around the drill) and blurry spikes in the temporal slices (b), (c), (d)). Oh et. al [17] induce flickering motion (seen as spikes in the temporal slice (e)) and blurry distortions in some frames (visible in the frame (e)). The proposed networks ( (f) D1 and (g) D2) produce better magnification with fewer distortions.