normals of the polydata are messed up after transform

[Hanxin_Hua]

   vtkNew<vtkTransform> transform;
transform->PostMultiply();
transform->Scale(1.0, vtkMath::Norm(axisY) / m_Thickness, 1.0);
transform->Concatenate(C);
transform->Update();

vtkNew<vtkTransformPolyDataFilter> transformPolyDataFilter;
transformPolyDataFilter->SetInputData(tubeFilter->GetOutput());
transformPolyDataFilter->SetTransform(transform);
transformPolyDataFilter->Update();

vtkNew<vtkTriangleFilter> triangle;
triangle->SetInputData(transformPolyDataFilter->GetOutput());
triangle->Update();

vtkNew<vtkPolyDataNormals> normalGenerator;
normalGenerator->SetInputData(triangle->GetOutput());
normalGenerator->ComputePointNormalsOn();
normalGenerator->ComputeCellNormalsOn();
normalGenerator->AutoOrientNormalsOn();
normalGenerator->NonManifoldTraversalOff();
normalGenerator->ConsistencyOff();
normalGenerator->Update();

Code is as above. So I create a tube and then apply transform to it (C is the matrix) and then display the normalGenerator->GetOutput()

But here you can see, one of the face is pointing inward. (Blue means normals are pointing outwards and pink means inwards)

Can anyone helps me with this?
Thank you!

[dgobbi (David Gobbi)]

It looks like one of the end faces is wound the wrong way. That would cause vtkPolyDataNormals to produce “inward” normals instead of “outward” normals.

I suspect that vtkTransformPolyDataFilter is not the problem. Instead, I think you might have discovered a bug in vtkTubeFilter: it winds one end cap correctly, but winds the other end cap inside-out.

If you remove vtkTransformPolyDataFilter, but keep vtkPolyDataNormals, do you get the same result? Try the following pipeline:

vtkTubeFilter -> vtkTriangleFilter -> vtkPolyDataNormals

[Hanxin_Hua]

If I remove vtkTransformPolyDataFilter, everything looks correct. And If I only remove the Concatenation, it also works. I also tried some combinations of rotation matrix. And some work but some don’t.
Also, I tried to use vtkCylinderSource as well. It works the same as tube. They both have issues.

Thank you!

[dgobbi (David Gobbi)]

Can you give a sample C matrix that doesn’t work?

[Hanxin_Hua]

For example, I just did this:
transform->RotateX(45);
transform->RotateY(60);
transform->RotateZ(80);

And the Tube’s axis is along z axis

[dgobbi (David Gobbi)]

I can confirm that the problem exists. In fact, for some rotations both of the end caps are incorrectly oriented.

But the problem goes away with this small change:

- normalGenerator->AutoOrientNormalsOn();
+ normalGenerator->AutoOrientNormalsOff();

This is very strange, because rotating the polydata should have absolutely no impact on this setting. This indicates either a bug in vtkTransformPolyDataFilter (which I doubt), or a bug in the AutoOrientNormals code in vtkPolyDataNormals (which seems more likely).

[Hanxin_Hua]

I tried that too.
If it is turned off, the whole object is pointing inwards. If I turn it on, there is this one face pointing inwards.
Is there a good solution for me to make them all pointing outwards? I can simply flip it but not sure if it is always the case

[Hanxin_Hua]

Yeah. In my cases, different rotations make different situations.

[dgobbi (David Gobbi)]

It seems the only reason that all normals would point inwards is if the m_Thickness in your code is negative.

[Hanxin_Hua]

Just checked, it is positive. The bug is still there even I comment out the scale

[dgobbi (David Gobbi)]

You’ll have to show how to reproduce the “all pointing inwards” situation, then.

[Hanxin_Hua]

lineSource->SetPoint1(0.0, 0.0, 1.0);
lineSource->SetPoint2(0.0, 0.0, -1.0);
lineSource->Update();

vtkNew<vtkTubeFilter> tubeFilter;
tubeFilter->SetInputData(lineSource->GetOutput());
tubeFilter->SetNumberOfSides(30);
tubeFilter->SetRadius(1.0);
tubeFilter->CappingOn();
tubeFilter->Update();

This is the tube creation
Then I just turned off the AutoOrientNormals

[Hanxin_Hua]

This is the C matrix:

-0.99830902570539681 0.057867716719294471 0.0055151207454731122 -4.1893909586657259 0.053899618858214371 0.95701416540486317 -0.28498897926301042 -0.17257266383903502 0.021769710197448333 0.28420980731869339 0.9585149269269575 6.0625220235805708
0 0 0 1

[dgobbi (David Gobbi)]

That’s not a rotation matrix, it has a determinant of -1 instead of +1. How was it computed?

[dgobbi (David Gobbi)]

In any case, you can probably fix the issue by using AutoOrientNormalsOff() and by adding the following to your pipeline (it doesn’t matter where, any place after vtkTriangleFilter will work).

// check if the matrix turns the data inside-out
bool flipped = (C->Determinant() < 0.0);

vtkNew<vtkReverseSense> reverse;
reverse->SetInputData(xxx->GetOutput());
reverse->SetReverseNormals(flipped);
reverse->SetReverseCells(flipped);
reverse->Update();

[Hanxin_Hua]

Yeah! It works! I might screw up the coordinates. But thank you!

[Hanxin_Hua]

Hi David, just some follow ups. I used an OBB Tree for an object, and then use (max, mid, min) as the new axes for the new coordinates. And I found out that (max, mid, min) do not make a right-handed coordinates. This is the issue causing the negative Determinant

[dgobbi (David Gobbi)]

Ah. When this happens, one option is to use a different corner to define the bounding box, e.g.

   if (vtkMath::Determinant3x3(max, mid, min) < 0.0)
   {
     for (int i = 0; i < 3; ++i)
     {
       corner[i] += size[2]*min[i];
       min[i] = -min[i];
     }
   }

[Hanxin_Hua]

Thank you!