Since this is a graphics hardware conference, I would like to
add that it seems feasible to support the core of this algorithm
natively in graphics hardware.
One might imagine an API that allows the user to first download the silhouette map to the hardware and then allows them to sample into a texture based on the silhouette map.
This would allow the examples I showed today to be implemented without a single fragment program! This would obviously mean a tremendous speedup for these applications.
The same hardware could also be used to accelerate the rendering stage of the silhouette map shadow algorithm since it essentially does the same thing.