What kind of surface doesn't fit this computational model? The elephant one for example, how to map the region which is occluded by the elephant's nose to the PVA surface? Anyway, really nice work. Looking forward to reading the paper and learn more!
The failure modes listed in the paper are models with a lot of concavities or self-occlusions. Looking at the video, it looks like the elephant cup succeeds because of surface tension properties of the water, which let the surface on each side of the trunk "snap together" when the trunk begins to be completely immersed. It's that wrapping or stretching that lets it be more than a simple light projection, which would fail due to self-occlusion.