Datasets:
How are the meshes in the PhysicalAI-Robotics-NuRec dataset obtained?
Hi, thanks for releasing the PhysicalAI-Robotics-NuRec dataset!
From what I understand, 3D-GRUT provides methods to generate Gaussian Splatting (3DGS) representations. However, in this dataset, I also see that meshes are provided along with the Gaussian models.
My questions are:
How were these meshes obtained?
Were they reconstructed from Gaussian Splatting?
Or were they directly captured / reconstructed from other 3D scanning methods (like depth fusion, photogrammetry, etc.)?
Is there a recommended workflow if we want to generate similar meshes for our own 3DGS data?
Understanding this would really help for downstream tasks such as physics simulation in Isaac-Sim, where a mesh representation is required.
Thanks a lot for your clarification!
Hi @jzhu98
Mesh is created from nvblox (https://github.com/nvidia-isaac/nvblox) by fusing depth images estimated from stereo RGB cameras.
Depth are estimated by FoundationStereo (https://nvlabs.github.io/FoundationStereo/)
We estimated image poses with cuSFM (to be released) and used the same pose both NuRec and Nvblox.
That's how mesh and NuRec are aligned so you can use them at the same time.
There will be a tutorial published in Oct about how to reproduce the results we generated, stay tuned. Cheers!
Wow! Thanks a lot
@willdzeng
for the detailed explanation!
I’ll definitely keep an eye out for the upcoming tutorial. Also looking forward to experimenting with the full pipeline!
Great post , thank you @willdzeng , look forward about this topic. Both 3d scene and mesh alignment is so great
@jzhu98 , @wayneji I am happy to announce that we recently released our reference workflow for how to create NuRec assets, please take a look. https://docs.nvidia.com/nurec/robotics/neural_reconstruction_stereo.html#
@jzhu98 , @wayneji I am happy to announce that we recently released our reference workflow for how to create NuRec assets, please take a look. https://docs.nvidia.com/nurec/robotics/neural_reconstruction_stereo.html#
Thank you , that's awesome
@jzhu98 , @wayneji I am happy to announce that we recently released our reference workflow for how to create NuRec assets, please take a look. https://docs.nvidia.com/nurec/robotics/neural_reconstruction_stereo.html#
Wow! Thanks! This really helps a lot.