Hacker News new | ask | show | jobs
by yorwba 3613 days ago
Not one of the authors, but I had a look at their paper http://www.interactivedynamicvideo.com/ISMB_Davis_2015.pdf

They use optical flow to get displacement values for each pixel. This works well when you don't have very large movements between frames, no sudden illumination changes and no rotations.

Car commercials will probably have all of those and on top of that be highly edited (e.g. you have to deal with cuts in the video)

You could try to use some other approach for motion estimation, like identifying and tracking salient points, but that would require quite a lot of work on top of what this demo shows.