Meshing random 5x5x5 fields, quickly found an example where standard MC33 is non-manifold (took around 5000 iterations). This happens when you get two marching cubes cases with tunnels adjacent to each other.
Video shows:
* upper-left: isosurface
* upper-right: Lopes method (manifold but complex and many triangles)
* lower-left: standard MC33 (not manifold in this case, non-manifold edges shown in red)
* lower-right: method I'm working on (manifold, only a few more triangles)
@sjb3d That looks fantastic. Is your approach MC based or is it something new?
@joeldevahl Thanks, yep it is still marching cubes, same classification of cases as MC33 but different triangulations.
It can add up to 4 interior vertices instead of just 1, and places these at similar feature points to the method by Lopes et al.
@sjb3d @joeldevahl nice! I would have guessed that after all these years MCs are a solved problem, but looks like they need you :)
@aras @joeldevahl finding an implementation that is both topologically correct (matches trilinear filtering) and manifold has proved elusive, hence this rabbit hole :)
@sjb3d @aras @joeldevahl have you looked at dual contouring? It's more costly, with finding tangentials, but the result is as good as it gets.
@dpwiz in my experience dual contouring does not match topology or generate manifold surfaces, so isn't what I'm looking for!
Dual methods are great for warping the triangulation around features though, but I plan to go in the direction of dual marching cubes.