geomdl approximate_surface not working as expected

Question

I am trying to approximate a image with a b-spline surface. The geomdl library has a function called "approximate_surface" to do this. When I use this with an image generated from a nice smooth functions such as f(x)=x^2, things work as expected. But when I use my actual image the control points seem to be all over the place. The b-spline surface is nothing like the input surface. Why is this?

I made an example using a small image:

img = np.array([
 [   0. ,    0.  , 590.2 ,   0.   ,  0.   ,  0. ],
 [   0. , 1115.8 , 1704. , 1724.2 ,   0.  ,   0. ],
 [   0. , 1390.6 , 1810. , 1877.8 ,1309.4 ,   0. ],
 [   0. , 1274.  , 1735. , 1720.  ,1249.8 ,   0. ],
 [   0. , 1000.  , 1345. , 1345.  ,1168.  , 348.2],
 [   0. ,  734.  , 970.  ,1105.   ,895.8  ,  0. ],
 [   0. ,  627.  , 805.  , 928.8  ,  0.   ,  0. ],
 [   0. ,  394.  , 635.2 , 627.2  ,  0.   ,  0. ]])
plt.imshow(img)
plt.show()

pts = []
h, w = img.shape
for i in range(h):
  for j in range(w):
    pts.append((float(i), float(j), float(img[i,j])))

surf = fitting.approximate_surface(pts,size_u=h,size_v=w,degree_u=1, degree_v=1)
ctrlpts = np.array(surf.ctrlpts)
print(ctrlpts[0:10])

I get the output

[[ 0.00000000e+00  0.00000000e+00  0.00000000e+00]
 [ 0.00000000e+00 -3.97922748e-01 -5.32705554e+02]
 [ 0.00000000e+00  2.72312015e+00  1.09167840e+03]
 [ 0.00000000e+00  4.07453113e+00 -3.98262686e+02]
 [ 0.00000000e+00  5.00000000e+00  0.00000000e+00]
 [ 5.97819029e+00  0.00000000e+00  0.00000000e+00]
 [ 5.97819029e+00 -3.97922748e-01 -1.16343388e+04]
 [ 5.97819029e+00  2.72312015e+00  1.93970077e+04]
 [ 5.97819029e+00  4.07453113e+00 -2.33379745e+04]
 [ 5.97819029e+00  5.00000000e+00  2.91199984e+03]]

plotted out:

vis_obj = vis.VisSurface(vis.VisConfig())
surf.vis = vis_obj
surf.render()

You can try out the code here in colab.

EDIT:

I ended up using scipy "scipy.interpolate.bisplrep" does behave as I expected.

The maintainers from geomdl don't seem to respond to the current open github issues. So if you have the same problem, my advice is to use the scipy functies.

"But when I use my actual image the control points seem to be all over the place, I have no idea why." is not a question. Please provide a specific technical question and the [Minimal Reproducible Example](https://stackoverflow.com/help/minimal-reproducible-example "Minimal Reproducible Example") that support your question — itprorh66, Jan 22 '22 at 16:09
@itprorh66 good point, I made it more specific. It is however as small as I can get it, if I make it even small there is not going to be a surface left anymore to get multiple cubic splines on. And if I try to generate the image from simple functions, it works. — zilleplus, Jan 22 '22 at 17:03

score 1 · Answer 1 · answered Jul 27 '22 at 14:40

I think the surface is working OK, but the viewer is not scaling sensibly. I have just re-run this code scaling the z direction by a factor of 1000 and it seems to look sensible.

import numpy as np
from geomdl import fitting
from geomdl.visualization import VisMPL as vis

img = np.array([
 [   0. ,    0.  , 590.2 ,   0.   ,  0.   ,  0. ],
 [   0. , 1115.8 , 1704. , 1724.2 ,   0.  ,   0. ],
 [   0. , 1390.6 , 1810. , 1877.8 ,1309.4 ,   0. ],
 [   0. , 1274.  , 1735. , 1720.  ,1249.8 ,   0. ],
 [   0. , 1000.  , 1345. , 1345.  ,1168.  , 348.2],
 [   0. ,  734.  , 970.  ,1105.   ,895.8  ,  0. ],
 [   0. ,  627.  , 805.  , 928.8  ,  0.   ,  0. ],
 [   0. ,  394.  , 635.2 , 627.2  ,  0.   ,  0. ]])

pts = []
h, w = img.shape
for i in range(h):
  for j in range(w):
    pts.append((float(i), float(j), float(img[i,j] / 1000)))

surf = fitting.approximate_surface(pts,size_u=h,size_v=w,degree_u=1, degree_v=1)

vis_obj = vis.VisSurface()
surf.vis = vis_obj
surf.render()

produces:

That's right on the money, after a while I figured this out as well. But totally forgot about this post. — zilleplus, Jul 29 '22 at 08:57

geomdl approximate_surface not working as expected

1 Answers1