Mesh Simplification is incomplete after one pass

[zalo]

@elalish It seems like there's room for additional simplification after one pass; perhaps topology should be simplified until no changes are made?

    general_offset = nonConvex.minkowski_sum(nonConvex.scale([0.5, 0.5, 0.5]))
    for i in range(5):
        save_mesh(general_offset, "general_offset_"+str(i), "stl")
        general_offset = general_offset.as_original()

Pass 0:

Pass 1:

Pass 2:

[elalish]

Yeah, that's probably a good idea for a stand-alone Simplify method - the internal one is aimed more at being fast dealing with what a single Boolean op creates.

[pca006132]

This is due to the order in which we perform edge collapses. We currently only do one pass over the list of halfedges in the order in the vector, but collapsing a halfedge with index i can cause halfedge j < i to be eligible for further simplification. Ideally we should have some kind of a worklist/set, when an edge is collapsed we check its neighbors to discover edges that can be further simplified. A worklist will be more efficient when the number of edges require collapsing is low (comparing with the total number of edges). I think this can combine with the optimization we mention in #679, i.e. only look for edge collapses among the newly-created verts in a Boolean, and will probably be faster than the current one we have.

[elalish]

Yeah, agreed. Currently EdgeCollapse is recursively checking neighboring edges for potential collapse, but I may be missing some cases.

[zalo]

Another interesting wrinkle... the slivers left over from some of the Minkowski operations appear to be cleaned up by the simplifications in asOriginal(), which also changes the Genus:

Before (Genus 4):

After calling as_original() 10 times (settled on Genus 5 after the first call):

Also worth noting the nearby triangle that is not getting simplified...

[pca006132]

Wonder what the black thing in your first photo is. Is this a bug in our boolean impl? And yeah, the lump of triangles not being simplified at the end is also interesting.

[elalish]

I would guess the black thing is a thin, disconnected mesh that's likely responsible for the genus being 1 low (an extra separate mesh will do that). On a second pass, I'm guessing it got simplified down to nothing.

[zalo]

Yep; that’s right! This is the same bug that was seen with Offset with the sliver meshes.

Simplify can often eliminate them (but not always).