Computer Vision News - February 2017

Lines 2-9 compute the approximate Riesz transform, as described in equation 2. Lines 10-12 compute the quadrature pair and the local phase. Note that the phase needs to be wrapped between –pi and +pi. Line 13 computes the local amplitude. Lines 14-16 compute the local phase difference between two successive images. This phase difference is filtered to remove all the non-dominant variation. Lastly, in line 18, this phase difference is used to amplify the movement (ref. here ) . Results: In terms of the video magnification quality, it is best seen in the glass video (if you didn’t watch it yet, you get here a second chance ) . Additional demos can be found on the research website . Run-time Improvement (the focus of this method): The above table compares Matlab implementations of phase-based motion magnification, the Riesz pyramid, and several variants of the complex steerable pyramid. For each sequence, the fastest phase-based method is highlighted (run- times, in seconds). Sum up: The research paper reviewed proposes a new representation, the Riesz pyramid, to replace the complex steerable pyramid, making real-time phase-based video magnification possible, without substantial loss of quality. The method may be relevant for more applications where sub-band phase is important, for instance stereo matching and phase-based optical flow. Computer Vision News Research 35 Research A new representation, the Riesz pyramid, to replace the complex steerable pyramid