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).