Edited by author 11-06-2003 07:23 PM
Regarding Neil's question 2 about why the background shape layers sum to 1, lumping the background together into one pdf is a functional simplification. Doing so allows the front elements of the "background" to be considered the same class with a common motion as predicted by the motion model, so the algorithm will not need to assign the predicted "background" motion to each occluding "background" layer. One of the primary assumptions is that, that the "background" has a common motion model. I can imagine that we model the layer visibility probability for all the layers just as was done for each foreground layer, and then assume that every other layer is the background layer or some combination of layers be assigned the background, but that would mean an additional step is required to decide whether or not each layer contain the common "background" motion and classify each as such.

I like the idea of layers for segmentation. Specifically, I like the idea of using a deformable appearance model to distinguish between multiple objects in images. I can imagine that there would be a tradeoff between finding a model to describe appearance deformations to keep objects a single layer and allowing objects to be divided into different layers based on objects lacking fit with that particular apperance model.