I remember reading this paper. ;) It is cool to think that a person only 30 pixels tall can provide sufficient information to determine the person's action (and then be able to mimic it with "do as I do"). This paper also reminds me a lot of the "Recognition of human gaits" paper that we looked at recently.

I do wonder about the skeletal transfer system. By using either the hand-marked joint locations or the 3D motion capture data, these are databases that are used outside the context of the scene. It seems to me that it would be better to build a training set using the actual video sequence itself. It may be that they decided not to do this because the resolution of a person is not high enough to determine the exact action. Does anyone think that using the "human skeleton represented as a kinematic chain" (from the "Recognition of human gaits" paper) would work here?

I like this paper. =) Using optical flow to get a motion descriptor isn't exactly intuitive, but once the idea is introduced, it is fairly straight forward.

Even though the authors say the main contribution of the paper is in the motion descriptor, I like their idea of "Do as I do" and "Do as I say". Maybe this can be used to creat personalized avatars in virtual reality applications?

At first I thought this paper was trying to recognize action just like the "Recognition of human gaits" paper, but seem that they are trying to do different things. The resolution of the figures used in this paper are probably too low to apply the technique of representing human skeleton as a kinematic chain.

I also like this paper. Comparing optical flows as a similarity measure is a pretty cool idea. One thing to be wary of (as Serge pointed out to me today) is that the background is relatively constant in most of their videos. If this were not so, I imagine there would be a lot more noise in their motion descriptors, which could lead to a breakdown of their classification mechanism.

This paper uses optical flows to extract the motion features from sequential frames, and compares different sequences using the correlation of these features. It is a simple method but works quite well. It compares the motion descriptors in both the spatial and temporal dimension and also solves the problem when the motion rates are different in two sequences. The synthesis part of "Do as I Do" and "Do as I Say" sounds fancy, I am looking forward to seeing the demonstration.

Very interesting paper, but it seems obviously failed to detect jump or similar actions unless they have some tricks in the stabilized proccess.

That's one of the weaknesses of this algorithm; that you have to experience it to be able to classify it. This means that unless the jumping, or some other random action was included in some sequence stored in a database already, this algorithm won't be able to classify it.

Stabilization is just one of the many assumptions the paper makes, I don't think it tries at all to do it...