search

PJLab-ADG / neuralsim

neuralsim: 3D surface reconstruction and simulation based on 3D neural rendering.
MIT License
582 stars 31 forks source link

How scalable is StreetSurf to larger-scale scenes #35

Open Geniussh opened 10 months ago

Geniussh commented 10 months ago

Thanks for the great work. The demos I have seen from the project website are all quite short (within 10 seconds) and travel only limited distance. Have you tried to train StreetSurf on larger scenes such as those that travel hundreds of kilometers? If so, will the current framework face any scalability issues? I am curious about why you haven't followed the same scene splitting approach as used in Block-NeRF, Mega-NeRF, and SUDS in order to readily scale to larger scenes.

I understand the distant-view branch might not be impacted because NeRF++ is internally robust to such large scenes with its inverted reparameterization. But how about the close-range branch that is based on NGP + NeuS? Will the hash-grid have enough capability to store features for larger scenes, or do you have to adjust N_{max}, T, and L (notations used in NGP paper) accordingly based on the size of the training sequence?

Looking forward to your answers!

amoghskanda commented 5 months ago

@Geniussh did you try rendering Streetsurf with differnet nvs paths?