anj1 / thinplatesplines.jl Goto Github PK

This small module implements thin-plate splines for Julia. Thin-plate splines are a very useful technique for representing smooth, continuous deformations in 2, 3, or N-dimensional space. They are often used in data interpolation, shape matching, and geometric design. Thin-plate splines allow you to specify a deformation on the entire space by specifying a set of 'control points' - their original and deformed positions. You can then 'extrapolate' the resulting deformation to the rest of the space, in a way that minimizes the bending energy over the entire space.

In this module, deformations are represented using the ThinPlateSpline type. Three functions are exposed:

1. tps_solve, which takes as input a set of control points and deformed control points, and a stiffness coefficient, and produces a deformation.
2. tps_deform, which takes as input as set of points and a deformation, and deforms the points according to the deformation.
3. tps_energy, a utility function which returns the bending energy of a thin-plate spline. This function is useful when you want to compare different deformations to each other.

Consider the following points (vertices of a triangle in 2D):

x1 = [0.0 1.0 
      1.0 0.0
      1.0 1.0]

And consider deforming them slightly:

x2 = [0.0 1.0
      1.1 0.0
      1.2 1.5]

Now calculate the TPS associated with this deformation (setting the stiffness coefficient to 1.0):

tps = tps_solve(x1,x2,1.0)

And deform another set of points using this deformation (note that one of the points is the same, and will be deformed in the same way):

x = [1.0 0.0
     2.0 2.0]
tps_deform(x,tps)

The output should be something like:

 1.1  5.55112e-17
 2.5  3.5

We can also calculate the bending energy of this deformation:

tps_energy(tps)

The result should be 0, which indicates that it is a fully affine transform.

How about a more general example? How about deforming continuous lines, not just points? tps_deform is general enough to be able to do that, without having to actually discretize the line as a lot of points. We can represent lines parametrically using linear functions of some variable t, with the TaylorSeries package. For example, let's represent the line in the plane given by x(t) = 2 + t, y(t) = 1 + t/2:

using TaylorSeries
t = Taylor1([0,1],15)
line = [2 + t, 1 + 0.5*t]'

We can deform this line just as we can with points:

deformed_line = tps_deform(line, tps)

This will look something like: [2.4 + 1.25 t + O(t¹⁶), 2.0 + 1.25 t + O(t¹⁶)]. In this case, since the transform is fully affine, the result is also a straight line, and so no terms higher than t exist. In general tps transforms, higher-order terms may exist. For instance:

x1 = [0.0 1.0
      1.0 0.0
      1.0 1.0
      1.0 2.0]
x2 = [0.0 1.0
      1.1 0.1
      1.2 1.5
      1.3 2.0]
tps_deform(line,tps_solve(x1,x2,1.0))

Gives a more complicated curve, with y(t) = 1.472 + 0.603 t - 0.009 t² + .... Note that this general idea can be extended to planes in 3D (or, indeed, higher-order dimensions) or even deformation of general geometric objects like spheres, etc.