NaN value in the rendered result

[quanhua92]

Describe the bug
I am working on a custom Renderer with Scene3DPytorch and ColoredTriMeshPytorch. However, I encountered a NaN in the output tensor of the scene.render method.
Here is the gist contains my code: link to code
The data for vertices and faces can be download at link to data

To Reproduce
Steps to reproduce the behavior:

1. Go to link to code
2. Download vertices.npz and faces.npz: link to data
3. Run the script nanproblem.py
4. See error:

• You can see that there are 387 nan values in the img tensor and the grad for the trans tensor is also nan
  print("nan counts = ", img[torch.isnan(img)].shape) # nan counts = torch.Size([387])
  print("trans grad", trans.grad) # trans grad tensor([nan, nan, nan])

Expected behavior
I expected to have a valid gradient in the trans.grad

Desktop (please complete the following information):

• OS: Windows 10
• Browser Chrome
• Version: 0.1.15 for NaN error.
  I have AssertionError on 0.1.17 at assert self.adjacencies.is_manifold
  Will update if have NaN error on 0.1.17

[quanhua92]

The vertices and faces are from a model that I am working on Open3D lib.
In DEODR 0.1.17, I can see that mesh.adjacencies.is_closed and mesh.adjacencies.is_manifold are both false.

I used some utility methods from Open3D lib on my mesh in Open3D and here are some results:

• open3d_mesh.is_watertight() => False
• open3d_mesh.is_vertex_manifold() => False
• open3d_mesh.is_edge_manifold() => True

Can you give me some ideas how to make sure mesh.adjacencies.is_closed and mesh.adjacencies.is_manifold are true?
Thanks

[martinResearch]

Hi,
I had a look at your mesh in meshlab and there are some non manifold edges. Mesh repairing is out of the scope of my library. how was your mesh produced ? could you use a method that would produce watertight surfaces (Poisson reconstruction for example)? Otherwise you could try to repair your mesh in meshlab.

[quanhua92]

Thanks. I used a neural network to output this mesh. I will find a way to produce watertight surfaces.

[martinResearch]

I have PR in progress that adds a method to extract the largest manifold subpart of a mesh using linear programming #51. However the best approach would be to have an algorithm that generates manifold surfaces in the first place