search

GGGHSL / GraphDreamer

[CVPR'24] GraphDreamer: a novel framework of generating compositional 3D scenes from scene graphs.
https://graphdreamer.github.io/
MIT License
160 stars 1 forks source link

About individualized object fields encoding in sec4.2 #8

Closed lxl1115 closed 5 months ago

lxl1115 commented 5 months ago

Thanks for amazing work! I would like to ask, during the optimization process of the implicit field, are both multi-resolution hash grid encoding(from Instant NGP) and the shared shallow MLPs(for SDF and color prediction) training simultaneously? And how are they initialized?

GGGHSL commented 5 months ago

Hello, thanks for asking. For the first question, yes, both the encoding networks and the MLPs are trained simultaneously. The hash encoders are randomly initialized (according to a seed), and the MLPs are using Kaiming-uniform distribution.