Hi Bruno,

If we use a multipolygon concept for the polyhedron, for a slightly non-convex polyhedron, the only way to compute its volume is by tetrahedral decomposition. If the tetrahedral decomposition fails for a multipolygon polyhedron, we will see inverted tetrahedrons as parts of the decomposition (negative mixed product volume). This might be our test per choice, to be used with the metafunction you mentioned. Of course, there is another way to compute the volume, which may be faster: analytical formula, but this requires strict convexity of the polyhedron. We can probably switch between those options somehow.

I'll start working with P1 as you suggested, since there is a lot to try out. When I read about intersection algorithms for polyhedra, often the complexity O is mentioned, but the real question is how much the intersection can be sped up for a naive intersection (O(m*n)) with collision detection for the polygons of the polyhedron (and or Axis Aligned Bounding Box [AABB] tests).

I'll try to get the P1, collision detection and AABB tests up and  running. Then we can apply the naive O(m*n) intersection like this:
 
P1 intersect (P1 X , P1 Y)
    P1 result
    if AABB intersect (X, Y):
        if collide (X, Y):
            for all polygons X (pX):
                for all polygons Y (pY):
                    if collide (pX, pY):
                       pResult = intersect (pX, pY)
                       append pResult to result
    return result

P2 which is graph based will wait for a while, but I would like to do that as well. The question is for reasonable number of polygons (I am dealing with polyhedra that have up to say 20 polygons), which algorithm is faster, the Preparata/Muller or the naive intersection with collision detection (and/or AABB test).

I'll start working with P1 as you suggested and see where it leads. As soon as I have something to show or to ask, I'll let you know. :)

Best regards,
Tomislav

On 05/22/2013 11:11 AM, Bruno Lalande wrote:
Hi Tomislav,

Good point, I hadn't realized you were merely using the existing multipolygon concept. I think it makes sense. Moving forward we should also propose validation functions to check it's actually is a polyhedron (as a polygon soup can be anything) but this can be done later.

Implementing the Separating Axis Collision test is one of the first thing I wanted to do when started to handle real £D stuff so I'm glad you're looking into this. The algorithm indeed relies on the polyhedron being convex. Which essentially means that our concepts are going to need a way to express that (i.e. to say "I am convex" so optimal algorithms like this one can be selected). Long ago I simply thought there would be a Polyheron and a ConvexPolyhedron concept but things are much more complicated that there are tons of ways to represent polyhedrons (not just the ones you described), all valid in some situations. I can see 2 solutions:
- ConvexPolyhedron could be a concept on its own, which a model can fulfill in addition to another one (e.g. it can be a AdjacencyGraph and a ConvexPolyhedron) - however this require changes to our current design I think.
- We could have a is_convex<Geometry> metafunction that algorithms would call when they're interested in whether a geometry is convex (could be used also with polygon and plenty other stuff - could be hardcoded to true for concepts like box).

So all in all, starting with your P1 is a valid approach. It terms of the IO stuff you've written it looks right since the format seems to be representing exactly that (a set of polygons without any strong guarantee about their spatial disposition). And you can try to work on other algorithms with this concept as well to see what it gives.

Regards
Bruno



_______________________________________________
Geometry mailing list
Geometry@lists.boost.org
http://lists.boost.org/mailman/listinfo.cgi/geometry