Richard Baraniuk in his presentation entitled Multiscale Geometric Analysis (the real player version of this presentation is here) makes a very good presentation of how compressed sensing works. What is interesting in this presentation are the questions of the audience at the end: They focus on the engineering of the mask/camera but in fact the new element is the mathematics (and that element is not a new wavelet basis). The progressivity aspect of the reconstruction is important and it is, in my view, shown very well in the recent bayesian paper on this. So in order to build a new camera ones needs a DMD controlled board (at 10K$), one pixel and you're set. The underlying reason why this concept is so important is when it brings large advantage over current solutions. For instance, movement detection. In traditional approach one needs to do some type of optic flow computation between two frames to evaluate changes.
With compressed sensing, since events shown on a camera are highly correlated to each other over time, Baraniuk shows that instead of having to deal/compute with 36 millions bits per second (for example using a traditional optic flow solution) one only needs to deal with 24 bits per second to evaluate changes. Another case in point is shown in the presentation on Intelligent Motion Detection Using Compressed Sensing by Heather Johnston, Siddharth Gupta, Grant Lee, Veena Padmanabhan.
No comments:
Post a Comment