I thought the feature classification part of this paper was really interesting. My question is that it seems like although they say that it would be a good preprocessing step for structure from motion and related techniques, it seems like they'd need a sequence of images to get their image trajectory for the classification - is it possible to do something like this in real time?