Computer Vision News - February 2017

34 Computer Vision News Research Research First, each frame in the video is decomposed into multiple sub-bands, each of which corresponds to a different spatial scale. For every two successive frames, the Riesz transform is performed, producing R 1 , R 2 . Next, the quadrature pair Q and the phase difference (φ t-1 - φ t ) are computed to form R 1 , R 2 (90 degrees phase-shifted) with respect to the dominant orientation (θ). Changes not along the dominant orientation (θ) are filtered out by an amplitude-weighted filter. And finally, the motion magnified sub-band is produced. Below is the pseudocode implementing motion processing with the Riesz transform for a given sub-band: