Question about the implementation of octree about nglod HOT 3 CLOSED

Thanks for your interest in our work.

In the paper, at training time we similarly initialize with a dense buffer, and make it sparse at inference time. If you know or can predict which octree cells will be occupied, however, there's no reason to not be able to use sparsity to your advantage at training time. We plan to showcase some of these use-cases, but they probably won't be ready for the first release (which will hopefully come very soon).

If you wanted to implement something like that yourself, though, all the code you'll need to implement that will be there in our first release.

from nglod.

To be more clear, I want to use sparse voxel tree in both reconstruction and rendering, which requires a sparse storage structure during the whole training pipeline.

from nglod.

It took some time but we finally pushed the sparse training & rendering code in Python: 9d447ee

Hopefully this can be useful for everyone!

from nglod.