Multivariate interpolation with RBF

1 Introduction

Making a grid flow through guide points using multivariate interpolation with the Python module Numpy / Scipy.

2 Initialization

The f@height attribute is going to be interpolated based on common X and Z positions between the grid and the guide points.

// Grid
f@px = v@P.x;
f@pz = v@P.z;

f@height = 0.0;

// Guide points
f@px = v@P.x;
f@pz = v@P.z;

f@height = v@P.y;

3 Interpolation

Python's NumPy/SciPy library offers multivariate interpolation. It's making sure the grid will always flow smoothly through our guide points even if they are positioned irregularily.

There are various interpolation methodes provided: multiquadric, inverse multiquadric, gaussian, linear, cubic, quintic and thinplate.

import numpy as np
from scipy.interpolate import griddata
import scipy.interpolate as interp

node = hou.pwd()
geo1 = node.geometry()
inputs = node.inputs()
geo2 = inputs[1].geometry()

method_nr = node.evalParm('method')
method_names = 'multiquadric,inverse_multiquadric,gaussian,linear,cubic,quintic,thin_plate'.split(',')
method_str = method_names[method_nr]

grid_x = np.array(geo1.pointFloatAttribValues('px'))
grid_z = np.array(geo1.pointFloatAttribValues('pz'))
pos_x = np.array(geo2.pointFloatAttribValues('px'))
pos_z = np.array(geo2.pointFloatAttribValues('pz'))
height = np.array(geo2.pointFloatAttribValues('height'))

rbf_height = interp.Rbf(pos_x, pos_z, height, function=method_str)
smooth_rbf_height = rbf_height(grid_x, grid_z)

geo1.setPointFloatAttribValuesFromString("height", smooth_rbf_height.astype(np.float32))

v@P.y = f@height;

4 Conclusion

The radial basis function (RBF) in Scipy.interpolate does an excellent job at interpolating between irregular points. RBF looks smoother than what we´d get from the Attribute Transfer SOP, too.

5 Contribution

Adrian Pan contributed a corresponding example file that works directly with heightfields.

6 VEX version

rbf.hip contains a simplified VEX version of multivariate interpolation using radial basis functions. The screenshot compares the attribute transfer node's result (left) with RBF.

int mode = chi('mode');
float k = chf('radius');

float w    = 0.0;
vector clr = 0.0;

for(int i = 0; i < npoints(1); i++){
    vector pos = point(1, 'P', i);
    float d = distance2(v@P, pos);
    vector clr_pt = point(1, 'Cd', i);
    
    float weight = 0.0;
    if     (mode == 0) weight = 1.0 / sqrt(d);                // inverse multiquadratic
    else if(mode == 1) weight = pow(1.0 + (d * d), -0.5) / d; // multiquadric
    else if(mode == 2) weight = exp(-d / (2.0 * k * k));      // gaussian
    else if(mode == 3) weight = 1.0 / (d * d * d);            // cubic
    w += weight; 
    
    clr += clr_pt * weight;
}

v@Cd = clr / w;

1 Introduction

2 Initialization

3 Interpolation

4 Conclusion

5 Contribution

6 VEX version

Downloads

Related articles

Neural Network SDFs

Batch processing files with Python

Custom clipping tools using Python verbs

Convert meshes to vector graphics

Machine learning tutorials for Houdini

Attribute to match across inputs