From meshes to deformation fields

In this tutorial, we show how the deformation fields that relate two meshes can be computed and visualized.

The following resources from our online course may provide some helpful context for this tutorial:

  • Modelling Shape Deformations (Video)
Preparation

As in the previous tutorials, we start by importing some commonly used objects and initializing the system.

import scalismo.geometry._
import scalismo.common._
import scalismo.ui.api._
import scalismo.registration.Transformation

scalismo.initialize()
implicit val rng = scalismo.utils.Random(42)

val ui = ScalismoUI()

We will also load three meshes and visualize them in Scalismo-ui.

import scalismo.io.MeshIO

val dsGroup = ui.createGroup("datasets")

val meshFiles = new java.io.File("datasets/testFaces/").listFiles.take(3)
val (meshes, meshViews) = meshFiles.map(meshFile => {
  val mesh = MeshIO.readMesh(meshFile).get 
  val meshView = ui.show(dsGroup, mesh, "mesh")
  (mesh, meshView) // return a tuple of the mesh and the associated view
}) .unzip // take the tuples apart, to get a sequence of meshes and one of meshViews 

Representing meshes as deformations

In the following we show how we can represent a mesh as a reference mesh plus a deformation field. This is possible because the meshes are all in correspondence; I.e. they all have the same number of points and points with the same id in the meshes represent the same point/region in the mesh.

Let’s say face_0, is the reference mesh:

val reference = meshes(0) // face_0 is our reference

Now any mesh, which is in correspondence with this reference, can be represented as a deformation field. The deformation field is defined on this reference mesh; I.e. the points of the reference mesh are the domain on which the deformation field is defined.

The deformations can be computed by taking the difference between the corresponding point of the mesh and the reference:

val deformations : IndexedSeq[EuclideanVector[_3D]] = reference.pointSet.pointIds.map {
  id =>  meshes(1).pointSet.point(id) - reference.pointSet.point(id)
}.toIndexedSeq

From these deformations, we can then create a DiscreteVectorField:

val deformationField = DiscreteField[_3D, UnstructuredPointsDomain[_3D], EuclideanVector[_3D]](reference.pointSet, deformations)

Similar to discrete scalar images, a Discrete Vector Field is defined over a discrete domain. In contrast to images, the domain does not need to be structured (a grid for example) and can be any arbitrary finite set of points. In the above example code, we defined the domain to be the reference mesh points, which is of type UnstructuredPointsDomain[_3D], as we can easily check:

val refDomain : UnstructuredPointsDomain[_3D] = reference.pointSet
// refDomain: UnstructuredPointsDomain[_3D] = scalismo.common.UnstructuredPointsDomain3D@e36f4d21
deformationField.domain == refDomain
// res1: Boolean = true

As for images, the deformation vector associated with a particular point id in a DiscreteVectorField can be retrieved via its point id:

deformationField(PointId(0))
// res2: EuclideanVector[_3D] = EuclideanVector3D(
//   -0.031402587890625,
//   -0.24579620361328125,
//   4.780601501464844
// )

We can visualize this deformation field in Scalismo-ui using the usual show command:

val deformationFieldView = ui.show(dsGroup, deformationField, "deformations")

We can see that the deformation vectors indeed point from the reference to face_1. To see the effect better we need to remove face2 from the ui, make the reference transparent

meshViews(2).remove()
meshViews(0).opacity = 0.3

Exercise: generate the rest of the deformation fields that represent the rest of the faces in the dataset and display them.

Deformation fields over continuous domains:

The deformation field that we computed above is discrete as it is defined only over the mesh points. Since the real-world objects that we model are continuous, and the discretization of our meshes is rather arbitrary, this is not ideal. In Scalismo we usually prefer to work with continuous domains. Whenever we have an object in Scalismo, which is defined on a discrete domain, we can obtain a continuous representation, by means of interpolation.

To turn our deformation field into a continuous deformation field, we need to define an Interpolator and call the interpolate method:

val interpolator = NearestNeighborInterpolator[_3D, EuclideanVector[_3D]]()
val continuousDeformationField : Field[_3D, EuclideanVector[_3D]] = deformationField.interpolate(interpolator)

As we do not know much about the structure of the points that define the mesh, we use a NearestNeighborInterpolator, which means that for every point on which we want to evaluate the deformation, the nearest point on the mesh is found and returned.

The resulting deformation field is now defined over the entire real space and can be evaluated at any 3D Point:

continuousDeformationField(Point(-100,-100,-100))
// res5: EuclideanVector[_3D] = EuclideanVector3D(
//   7.967502593994141,
//   1.7736968994140625,
//   -6.764269828796387
// )

Remark: This approach is general: Any discrete object in Scalismo can be interpolated. For more structured domains, such as the DiscreteImageDomain, we can use more sophisticated interpolation schemes, such as linear or b-spline interpolation.

The mean deformation and the mean mesh

Given a set of meshes, we are often interested to compute mesh that represents the mean shape. This is equivalent to computing the mean deformation , and to apply this deformation to them mean mesh.

To compute the mean deformation, we compute for each point in our mesh the sample mean of the deformations at this point in the deformation fields:


val nMeshes = meshes.length

val meanDeformations = reference.pointSet.pointIds.map( id => {

  var meanDeformationForId = EuclideanVector(0, 0, 0)

  val meanDeformations = meshes.foreach (mesh => { // loop through meshes
    val deformationAtId = mesh.pointSet.point(id) - reference.pointSet.point(id)
    meanDeformationForId += deformationAtId * (1.0 / nMeshes)
  })

  meanDeformationForId
})

val meanDeformationField = DiscreteField[_3D, UnstructuredPointsDomain[_3D], EuclideanVector[_3D]](
  reference.pointSet, 
  meanDeformations.toIndexedSeq
)

We can now apply the deformation to every point of the reference mesh, to obtain the mean mesh. To do this, the easiest way is to first genenerate a transformation from the deformation field, which we can use to map every point of the reference to its mean:

val continuousMeanDeformationField = meanDeformationField.interpolate(interpolator)

val meanTransformation = Transformation((pt : Point[_3D]) => pt + continuousMeanDeformationField(pt))    

To obtain the mean mesh, we simply apply this transformation to the reference mesh:

val meanMesh = reference.transform(meanTransformation)

Finally, we display the result:

ui.show(dsGroup, meanMesh, "mean mesh")