Edited by author 10-30-2003 10:39 PM
Optical flow is relatively appearance invariant. The implications of using it to recognize certain motions is very interesting. I hope there can be some extensive flow models of various human motions, creating a library with which to perform recognition tasks.

The work reminds me of the randomized looping of video sequences, e.g. taking a sequence of a candle flame and looping it over and over again at random places. In it, they also used the similarity matrix but on full image frames.

This paper claims that the novel contribution is the motion descriptor. There appears to be however so much other work on the side for it. It appears to me that this person spent a lot of time on presenting its promise: skelton transfer, action synthesis, figure correction, "action database", if these things were indeed not research contributions themselves as the author seems to imply. The action synthesis one certainly doesn't sound too new; candle flame looping uses a similar concept confusion matrix concept. I guess what i'm saying is a follow up paper is called for on all the "promise" made in the paper.