TensorToMesh: Node to convert a position and vertexIds tensors into a mesh.

[lucienfostier]

• GafferML: Add support for converting tensors into a mesh

Checklist

• I have read the contribution guidelines.
• I have updated the documentation, if applicable.
• I have tested my change(s) in the test suite, and added new test cases where necessary.
• My code follows the Gaffer project's prevailing coding style and conventions.

[johnhaddon]

Nice one, thanks Lucien! I've made a few comments inline but it's all pretty minor.

One thing I'm wondering is if different mesh-generating models in the wild all use this same format for defining the mesh data, or if there is lots of variation as there is for image data? Does this node just handle data from one specific model, or does it stand a good chance of being general purpose?

Cheers...
John

[johnhaddon]

We need documentation for this node and the plugs - the lack is causing a test failure at present.

[lucienfostier]

addressed here 9550acb

[johnhaddon]

Thanks for the plug docs - we just need a description for the node itself now.

[lucienfostier]

Nice one, thanks Lucien! I've made a few comments inline but it's all pretty minor.

One thing I'm wondering is if different mesh-generating models in the wild all use this same format for defining the mesh data, or if there is lots of variation as there is for image data? Does this node just handle data from one specific model, or does it stand a good chance of being general purpose?

Cheers... John

Honestly, I don't know...SMPL is the only mesh generation model I've been experimenting with and the one we plan to use soon at IE.

To avoid any bad surprise I added extra validation of the input tensors but the node might need some change in the future depending on further experiment with other models.

That being said, I'd like to think that it's a fairly reasonable assumption to make that a mesh should be a collection of 3d point coordinates and a list of vertexIds.

Do you have any suggestion to make this more generic or should we wait until we use different ML model generating or processing mesh to make adjustment?

After all I think GafferML is still in "experimental" so that's the time we can gather knowledge and break stuff...

[boberfly]

I've not tried this one yet but I was going to give this one a go eventually with GafferML - it might add some more insight to what kind of data it outputs:
https://github.com/microsoft/TRELLIS

[johnhaddon]

Honestly, I don't know...That being said, I'd like to think that it's a fairly reasonable assumption to make that a mesh should be a collection of 3d point coordinates and a list of vertexIds.

Do you have any suggestion to make this more generic or should we wait until we use different ML model generating or processing mesh to make adjustment?

I don't have any alternative ideas, and in the absence of other use cases I think this is fine - I was just wondering if you might know about other potential representations. I did take a brief look at TRELLIS and found some evidence that it's using the same layout.

Here is an optional commit that renames everything to match Gaffer's convention:

I think I do have a mild personal preference for the updated names, but will leave the choice to you.

Could you address the documentation for the node itself and squash everything down ready for merging please?