Phase-based motion magnification
Phase-Based Video Motion Processing — Wadhwa, Rubinstein, Durand, Freeman, MIT CSAIL, SIGGRAPH 2013
Definition
Phase-based motion magnification is the successor to Eulerian motion magnification (2012). It uses complex-valued steerable pyramids to encode motion as phase variation in spatial-frequency coefficients, then temporally filters the phase to amplify (or attenuate) imperceptible motions with significantly less noise amplification than the original Eulerian method.
Verbatim from the source
“Phase variations of the coefficients...over time correspond to motion, and can be temporally processed and amplified to reveal imperceptible motions, or attenuated to remove distracting changes.”
Technical detail
Compared with the original Eulerian intensity-based method (Wu et al., SIGGRAPH 2012), the phase-based approach has a much higher noise margin — it amplifies motion without amplifying photon-shot noise — and supports larger motions before the linear approximation breaks down.
The complex steerable pyramid is computationally heavier than the Laplacian pyramid used in the original Eulerian method. Real-time implementation on a phone GPU requires the entire pipeline to run on shaders rather than CPU. Apple Metal compute shaders are particularly well-suited for the per-pixel-per-band operations.
Where it appears
- Slick Engineering Research Used here
Phase-based motion magnification is the more capable of the two MIT methods Slick Engineering reverse-engineered into a real-time GPU-accelerated iOS implementation. The case study covers both.