I am looking for ideas for algorithms and data structures for representing volumetric objects. I am working on a sculpting system, like sculptrix or mudbox, and want to find a good implementation strategy.
I currently have a very nice dynamic halfedge mesh system to collapse/subdivide faces. It works very well and is incredibly fast, but since it is a surface algorithm, it is not easy to robustly change topology.
So I want to go back to the drawingboard and implement a proper volumetric system. My first idea was some kind of octtree representation for the volume and marching cubes to polygonise it.
However, I have a few problems with this. First, marching cubes often produces small or thin triangles, something that is highly undesirable (reason why later). Second, I want to polygonise the volume only in the area of editing, and at different levels of detail. For example, I may want a low res sphere, but with a few tiny high res bumps. I can easily get that kind of subdivision behaviour with my current surface based sustem, but I can't envision how I could do it robustly with marching cubes.
Another problem is that the actual trianglular mesh is further subdivided on the gpu for smooth surfaces, so I need neighbourhood information too. Again, I already have this with the current half-edge system, but with a volume polygonisation system, I imagine it taking a lot of extra processing to find the extra connectivity information. This is the reason thin triangles are bad.
So I have a lot of constraints, and I am asking this community for ideas or pertinent papers to read. I was thinking about surfacenets to avoid the small/thin triangle problem. Also, I have a feeling kd-trees may be better for storing multiresolution volumes since they seem more flexible then octtrees.
Anyway, any ideas/suggestions very welcome.