Geometric differentiation cannot optimize vertex positions at all.

[udemegane]

I'm not an English native, so I first apologize for the unnatural English.
I am using the "pathreparam" plugin (OptiX6.5) to test geometric differentiation. It transforms the base sphere object into another body.
However, it exhibits very unnatural behaviour.
The following is a video that attempts to transform a sphere into an ear.

• base object(sphere)

• ref object(ear)

I've tried multiple approaches, such as parameter tuning, regularization,
setting background objects and adjusting lighting etc.but I can't avoid mesh collapse.
Curiously, this collapse of the mesh changes shape each time you run your code. (Despite the exact same program)

Depending on the object to be transformed, the collapse of the mesh can be almost avoided.

• ref objects

I have no idea what is happening.

Thank you.

[arpit15]

Have you given the new version of the plugin a try?

[udemegane]

I tried the new version but it still doesn't work. If I make any progress, I will report it here.

[Speierers]

Hi @udemegane ,

Please have a look at the tutorials in the pathreparam-optix7 branch, this could be helpful.

How do you modify the mesh at every iteration? Do you let the optimizer move the vertices directly using gradient descent to are you implementing a intermediate step, like in the tuto? (e.g. constrained to displacement along the normals)?

[arpit15]

@Speierers I get errors when I run pathreparam ex. For ex, I ran invert_pose.py and I get following error

cuda_check(): runtime API error = 0002 "cudaErrorMemoryAllocation" in ../ext/enoki/src/cuda/horiz.cu:59.
terminate called after throwing an instance of 'std::runtime_error'
  what():  CUDABackend: referenced unknown variable 112575
Aborted (core dumped)

I have RTX 2080 with 11MiB RAM. When I build in debug mode, the gdb bt complains path-reparam.cpp:449

[Speierers]

It is possible that you are running out of memory on your GPU, the pathreparam integrator uses quite a bit of VRAM. Could you try to renderer with lower SPP maybe?

[udemegane]

@Speierers
I do not give direct constraints on individual vertices, such as constrained to displacement along the normals. But that's a good idea, so I'd like to put it into practice, but I couldn't find a description in the documentation that would limit the vertices. Am I overlooking something?

[arpit15]

@Speierers Yeah it turns out spp got me again. I reduced it to 16 and everything worked. Is there a rule of thumb for memory with pathreparam to decide on the image size and spp?

[udemegane]

I finally got the understanding about the pre render callback function. I also fully understood this issue by reading the pathreparam paper. All causes are simply that there were cases where the mesh was "not valid". Careful reading of the documentation tutorials will work around this issue, so I shut this out. Thank you for all your efforts!