Help implementing 3D SDF/UDF fitting

[lykius]

Hi,
This framework is really amazing, thanks a lot for sharing it.

I've been playing around with samples/mlp_learning_an_image.cu and I was wondering if it is possible to adapt it in order to fit a point cloud or a mesh by fitting its UDF/SDF.
I'm not really expert in CUDA programming, do you think it's doable with a reasonable effort? Do you have any suggestion?
I was thinking about computing coordinates and groundtruths in python and saving them in numpy files, then loading these numpy files inside the CUDA program (with cnpy). But after that, I'm not sure about the next steps. In the image case, you create a CUDA texture that seems quite specific for 2D images, how should I adapt it to 3D data?

Thanks in advance for any kind of help,
Luca

[Tom94]

Hi, absolutely! In our instant-ngp project, we've implemented a number of neural graphics primitives using this framework, one of which is indeed an SDF. You can run it and pass it a .obj-formatted mesh and it'll fit a corresponding SDF within a second or so + give you a live rendering preview.

Over in that project, see testbed_sdf.cu for the SDF-specific source code.

[lykius]

Hi @Tom94, thanks a lot for your answer!

I know instant-ngp, it's super awesome!! I was trying to use tiny-cuda-nn since I'm currently interested in a more general MLP formulation, without hash encoding. Maybe I can start from the SDF fitting in instant-ngp and remove the hash encoding. What do you think?

[Tom94]

instant-ngp's SDF code uses tiny-cuda-nn's NetworkWithInputEncoding, which can use any encoding that you'd like. All you need to do is change out the encoding in the .json configuration that you pass to the program via the --network argument. It loads <task>/base.json by default, where <task> is sdf in this case.

[lykius]

Ok I'll try that!
Thanks a lot, I really appreciate your help!

[Tom94]

Gladly!