Weird behavior in PoissonDiskSampler

[dyollb]

I am trying to use the PoissonDiskSampler to sample a triangle mesh with minimum sample distance distance:

    gc::surface::SimplePolygonMesh simpleMesh;
    for (const auto & t : qmesh.get_tris_buffer())
    {
        simpleMesh.polygons.push_back({t.inds[0], t.inds[1], t.inds[2]});
    }
    for (const auto & v : qmesh.get_verts_buffer())
    {
        simpleMesh.vertexCoordinates.push_back({v.x, v.y, v.z});
    }

    std::unique_ptr<gc::surface::ManifoldSurfaceMesh> mesh;
    std::unique_ptr<gc::surface::VertexPositionGeometry> geometry;
    // throws exception if input is non-manifold/invalid
    std::tie(mesh, geometry) = gc::surface::makeManifoldSurfaceMeshAndGeometry(simpleMesh.polygons, simpleMesh.vertexCoordinates);

    geometry->requireEdgeLengths();
    double meanEdgeLength = 0.;
    for (auto e : mesh->edges()) {
      meanEdgeLength += geometry->edgeLengths[e];
    }
    meanEdgeLength /= mesh->nEdges();
    geometry->unrequireEdgeLengths();

    gc::surface::PoissonDiskSampler sampler(*mesh, *geometry);
    auto samples = sampler.sample(distance / meanEdgeLength);

    std::vector<Vec3d> points;
    points.reserve(samples.size());
    for (auto& s: samples)
    {
        auto p = s.interpolate(geometry->vertexPositions);
        points.push_back({p.x, p.y, p.z});
    }

If I sample a (approximate) cube with edge length 40 (two triangles per cube face), I get following behavior:

distance=4	distance=2	distance=1
samples: 3	samples: 22	samples: 6559

Why is this happening? Am I doing something wrong?

[dyollb]

@nmwsharp Is this a bug or a user error?

Looking at the code I am wondering why the sqrt of the minDist is used to create new samples here:

geometry-central/src/surface/poisson_disk_sampler.cpp

Line 93 in 45321a7

double dist = std::sqrt(randomReal(rMinDist, 2. * rMinDist));

[nzfeng]

Hey @dyollb, thanks for making a pull request! I've been meaning to correct this issue for a while now. Totally agree it doesn't usually make too much sense for the min distance to be in terms of edge length.

And yeah, line 93 looks like a bug -- thank you for catching it!

I'll get back to you on your PR!