I was amazed at the various number of concepts and levels that were drawn from and put together to get this whole system to work. This resulted in a fairly dense paper that covered a broad spectrum of concepts. However, I did find the paper to be fairly well written even given this complexity.

I would have liked to have heard a bit more about some of the previous work and the work that was built upon to create this project. The paper seemed to somewhat dismiss this. Also, I think a few more examples of the various levels of the system working may have helped clarify how these different levels interact.

This is a very impressive paper. Sadly I couldn't fully understand the image model, even at the blob level. If you are presenting the paper tomorrow, would you please go through equation (2) , (3), and (4) slowly? Intuitively they seem to make sense, but the more I look at it the more I feel uncertain about their meaning.

I also can't understand what the meaning for equation (3). I can see that it is diretly derived from the relation between joint distribution and conditional distribution, but I don't know what the P(x,y|theta(k)) and the mixing coefficients stand for. So I'm expecting tomorrow's (maybe should be today?) representation.

Edited by author 11-20-2001 01:16 PM
I wasn't at all clear what is the relationship between the motion model and the mixture of blobs model. The last paragraph of the motion model section is particularly obscure - which term in (4) is he talking about and how does it define the probability?

I also liked the decomposition of human dynamics and the model that the paper has presented. As Dave said, the paper was interesting but was a little too dense. I hope that the talk will make some make the paper more concrete with some more examples.

I find the paper to be rather cool.
I have always wondered if tasks like this could be performed. Fundamentally what I like best is the analogy with speech recognition and the idea of not makin ghard choices at lower level and carrying the soft choices up the hierarchy. Which is very much inline with the ideas talked talk about while designing the Normalized Cut. This is a very elegant demonstration of the power of a Gestalt like approach.

Junwen: P(x,y|theta_k(t)) is essentially the probability
that the pixel x,y is a part of the kth block given the position and orientation (mean and moments, which is essentially the center of the blob and the principal axis if the ellipse best modelling it) and the position of the pixel itself (x,y).