How to sample instances of ShapeGroup?

[woAIxuexiSR]

Hi there,

I'm currently working on sampling shapes within my scene using the following code:

for shape in scene.shapes():
    ps = shape.sample_position(0, sample2, True)

However, when the iteration reaches an instance of a ShapeGroup, specifically:

Instance[
  shapegroup = ShapeGroup[
    name = "door_group",
    prim_count = 2956
  ]
  to_world = [[1, 0, 0, 0],
              [0, 1, 0, 0],
              [0, 0, 1, 0],
              [0, 0, 0, 1]],
]

I encounter the following error: RuntimeError: ​[Shape] Instance::sample_position(): not implemented!
I'm seeking guidance on how to properly sample shapes within a scene when encountering instances of ShapeGroups. Is there a recommended approach or workaround for this issue?

[njroussel]

Hi @woAIxuexiSR

A trivial, albeit inefficient, workaround would be to remove the instancing and just duplicate the shape in the scene description.

I don't there is any other option, this has not been implemented.
If you are interested in working on this, I don't believe that this would be a very difficult task. It should just be a matter of sampling a shape in the group, then sampling a position on that shape and finally applying the instance's transform to that position.

[woAIxuexiSR]

Thank for your help.

I want to dynamically apply a transformation to the shape. If I don't use an instance, I can't seem to find the 'to_world' attribute to modify when traversing through the scene parameters.

For example, I have an obj mesh in my scene.

<shape type="obj" id="Door">
    <string name="filename" value="models/Mesh006.obj" />
    <ref id="DoorBSDF" />
</shape>

I get this scene paramters:

Door.faces                                                     UInt              OBJMesh
Door.vertex_positions                                 ∂, D     Float             OBJMesh
Door.vertex_normals                                   ∂, D     Float             OBJMesh
Door.vertex_texcoords                                 ∂        Float             OBJMesh

Is there any alternative method other than applying the transformation matrix to each vertex and normal individually?

[njroussel]

Is there any alternative method other than applying the transformation matrix to each vertex and normal individually?

No. We have an example of how to do this, it's only a few lines: https://mitsuba.readthedocs.io/en/latest/src/inverse_rendering/object_pose_estimation.html#Optimizer-and-latent-variables.

Changing the vertex positions of a Mesh will make it recompute its normals by itself (it's an implementation of this method). These might not be exactly the normals you originally used. If this is an issue, you can apply a second separate update where you only update the vertex_normals.