Dolphine to Sphere Mesh about pytorch3d HOT 4 CLOSED

Hi @shersoni610! The purpose of all of our tutorials is purely instructional. The example of a deforming sphere serves as a way to highlight the various differentiable losses/regularizers we have implemented for 3D data. Our goal is to show how these ops can be defined and used when optimizing an objective with SGD. (Note that they're not limited to the purpose of the tutorial. In Mesh R-CNN, we use a similar set of losses/regularizers to learn how to fit an arbitrary topology (genus > 0) to a target shape. )

You might be correct in saying that classical approaches can also deform genus 0 shapes well. We don't claim that the tutorial is state-of-the-art in sphere deformation. Again, its purpose is purely instructional.

from pytorch3d.

from pytorch3d.

Hi @shersoni610! This is a valid question of course. Libraries provide nails and hammers and it's up to the user to find good use cases for them, if any. For me, it's obvious what we can start exploring with PyTorch3D which classical methods cannot offer. Though, a healthy research community is a diverse community, so I do appreciate and would strongly encourage these parallel efforts (ML vs. non ML, learning vs. non learning, etc.) to proceed. I will close the issue for now as it's not related to a bug/feature in the library, but I am happy to keep this discussion going!

from pytorch3d.

@gkioxari , I am considering trying 3D deep learning for my project, and trying to understand how PyTorch 3D relates/compares to NVIDIA's Kaolin. As this is relatively new, I have not found any comparisons online yet. Could you help explain, please?

from pytorch3d.