I constructed the Laplacian pyramid by forming a Gaussian pyramid and taking piecewise differences between each level and the upsampled version of the next level below, taking the last level of the gaussian pyramid as a base case. I then applied band-pass filter at every level of the Laplacian pyramid and added that back into the original signal with a weight (called alpha) depending on the pyramid level.
Results - Face Heartbeat
The video with and without magnification:
Here is the frequency spectrum of the "Y" component of the video, accross all Laplacian pyramid levels, before and after magnification (click for full version):
Notice the pre-existing spike at entry 10, which corresponds to about 1 Hz.
For reference, the alpha values used (from low to high spatial frequency layers): [30 22.5 15 7.5 0 0 0 0]
Results - Baby Heartbeat
The video with and without magnification:
Here is the frequency spectrum of the "Y" component of the video, accross all Laplacian pyramid levels, before and after magnification (click for full version):
Notice the pre-existing spikes around entries 34 and 36, which corresponds to about 2.5 Hz.
For reference, the alpha values used (from low to high spatial frequency layers): [100 100 2 0 0 0 0 0]
Results - Facial Expressions
My goal with this video was to magnify subtle facial twitches. It sort of works. It's mroe creepy than anything else. Observe before and after magnification:
For reference, the alpha values used (from low to high spatial frequency layers): [6 4.5 3 1.5 0.75]