Bugs in ops.mesh.check_sign() triggered under special circumstances

[Xiaoyao-Li]

Bugs in ops.mesh.check_sign() triggered under special circumstances

We found that if a point is in a sepcific axis, the check_sign is invalid for some primitive meshes.

from kaolin.ops.mesh import check_sign, index_vertices_by_faces
import trimesh as tm
import torch
# we test the check_sign function
x_points = torch.Tensor([[[1, 0, 0]]])
y_points = torch.Tensor([[[0, 1, 0]]])
z_points = torch.Tensor([[[0, 0, 1]]])
# e.g. When the test primitive is Sphere, check_sign is incorrect on x/y-axis
sphere_mesh = tm.primitives.Sphere(radius=2.0, subdivision=3)
points = z_points
print(f'trimesh contains() result: {sphere_mesh.contains(points.squeeze(0))}')
print(f'check_sign() result: {check_sign(torch.Tensor(sphere_mesh.vertices).unsqueeze(0), torch.Tensor(sphere_mesh.faces).long(), points)}')

Note that the cast a ray from the point is a really smart idea, and it is valid for most of points in our test, we don't know if this is an imperfection in the implementation.

[Caenorst]

Hi @Xiaoyao-Li , thank you for your interest, I'm exploring the issue. From what I'm seeing of the case the problem seems to be occuring from when the ray is intersecting a vertice or an edge.

Note: both cuda and cpu are failing, however returning different results.

[Xiaoyao-Li]

Thank you for this helpful info in your timely reply. We are benefitted from kaolin a lot and looking forward to a new version.

[Caenorst]

Hi @Xiaoyao-Li ,

Just merged the fix on GPU version, don't have enough time for the CPU version but I would happily merge an MR if someone wants to adapt the fix to CPU.

[Xiaoyao-Li]

Thank you for your hardworking again! Keep opening this issue for someone to fix on the CPU version.