This demo illustrates the dynamics of using belief propagation to
analyze transparency using our model. Our model prefers decompositions
with less edges and less corners. We represent a decomposition via its
gradient field. Thus there are two gradient fields, g1 and g2 (for the
two layers) and in BP we calculate the marginal probability at every
pixel for g1 and g2. In the discretization we use here, each gradient
can take on one of at most four possible values. When a pixel in g1
has a strong belief in a nonzero gradient, it turns red in the animation and when a pixel in g2 has a strong belief in a nonzero gradient, it turns green.