SSM with no triangulation - read-write inconsistency

[madsendennis]

There is some inconsistency to the StatisticalModelIO when reading and writing an SSM without the reference mesh having a triangulation (empty TriangleList).
A file can successfully be written. But when read, it fails due to the processing of missing items in "representer/cells".

Below is a small example which will show the error:

object ModelSaveLoadTest extends App {
  implicit val Random: Random = scalismo.utils.Random(1024)

  println("starting app...")
  scalismo.initialize()

  val zeroMean = VectorField(RealSpace[_3D], (_: Point[_3D]) => EuclideanVector.zeros[_3D])

  val cov: MatrixValuedPDKernel[_3D] = new MatrixValuedPDKernel[_3D]() {
    private val kernel = DiagonalKernel(GaussianKernel[_3D](1), 3)

    override protected def k(x: Point[_3D], y: Point[_3D]): DenseMatrix[Double] = {
      kernel(x, y)
    }
    override def outputDim = 3
    override def domain: Domain[_3D] = RealSpace[_3D]
  }

  val gp = GaussianProcess[_3D, EuclideanVector[_3D]](zeroMean, cov)

  val pointDomain = UnstructuredPointsDomain[_3D](IndexedSeq(Point3D(0, 0, 0), Point3D(1, 0, 0), Point3D(0, 1, 0)))
  val mesh = TriangleMesh3D(pointDomain, TriangleList(IndexedSeq()))

  val sampler = RandomMeshSampler3D(mesh, 10, 1024)
  val lowRankGP = LowRankGaussianProcess.approximateGPNystrom(gp, sampler, numBasisFunctions = 100)

  val ssm = StatisticalMeshModel(mesh, lowRankGP)

  val modelFile = new File("/tmp/dummy.h5")
  StatisticalModelIO.writeStatisticalMeshModel(ssm, modelFile)
  println("File written")
  StatisticalModelIO.readStatisticalMeshModel(modelFile).get
  println("File read")
}

Exception in thread "main" ncsa.hdf.hdf5lib.exceptions.HDF5Exception: ncsa.hdf.hdf5lib.exceptions.HDF5Exception: No data to read.

The error traces back to line 312 in the StatisticalModelIO.scala file.

[madsendennis]

Closing issue as the discussion/issue will continue in #426

[Andreas-Forster]

This issue goes futher than #426. #426 is needed to write and read back in UnstructuredPointsDomains. But fixing #426 will facilitate fixing this issue. It will remain to change StatismoDomainIO to allow the construction of the different domains also without cells. Currently this fails with a Throwable.

If it makes sense to write/read TriangleMeshes without cells is for me something we should agree on in the community. I have not a too strong opionion in either direction.

[madsendennis]

Thank you for looking more into this @Andreas-Forster. I'm okay with the solution as long as it is consistent with reading and writing. However, I lean more towards keeping the flexibility of writing a triangleMesh with no cells (and even no points) - since this is a valid triangleMesh in Scalismo, I don't see why we then shouldn't be able to read/write it.

[Andreas-Forster]

I would not block the additional flexibility, if we think it's the strategy in scalismo to allow that.

I think the discussion is, is a TriangleMesh with no cells, and maybe no points still a TriangleMesh? In case we say this is a valid TriangleMesh, even that it is maybe not a very usefull one, we should change the handling of the situation where no cells are read. Otherwise we should make the type TriangleMesh stricter.

@marcelluethi What do you think? Do you have any preference?

[marcelluethi]

Thanks for starting this discussion. We should allow for triangle meshes, where not all the vertices are connected with triangles, as this situation can occur if we clip with a predicate. If we allow some vertices to be not connected with others, I would find it consistent to allow all vertices have no triangles.

[Andreas-Forster]

Ok this sounds reasonable. I will then update the PR #427 to fix also this issue.

[Andreas-Forster]

I update the PR #427, so this issue should also be fixed when we merge the PR.