Hyperparameter suggestions for large scenes?

NVlabs / neuralangelo

Official implementation of "Neuralangelo: High-Fidelity Neural Surface Reconstruction" (CVPR 2023)

https://research.nvidia.com/labs/dir/neuralangelo/

Other

4.31k stars 387 forks source link

Hyperparameter suggestions for large scenes? #117

Closed hvkwak closed 11 months ago

hvkwak commented 11 months ago

Hi, thank you very much for your incredible work. I am trying to reconstruct a large scene of my custom dataset, where I used the hyperparameters of dict_size=22, dim=8 with all camera poses downward watching. Although this is a modular approach(which reconstructs only a part of the scene), the quality of the extracted mesh is not quite promising. Here's an example after the iteration of 300K:

I have got two questions:

May more iterations be helpful?
Any hyperparameter suggestions for better reconstruction of large scenes?

mli0603 commented 11 months ago

Hi @hvkwak

What is your batch size for the data?

hvkwak commented 11 months ago

Hi @hvkwak

What is your batch size for the data?

Hi @mli0603, oops! I kept it 1. Let me train this again with the batch_size of 16. Thank you!

iam-machine commented 11 months ago

@hvkwak Hi :) What are your results now after training with batch_size of 16?

hvkwak commented 11 months ago

@iam-machine @mli0603 Batch size was the key idea to reconstruct large scenes (more specifically 512 randomly sampled rays from all images per iteration), where coarse = 512, fine = 128 capture the geometry well. Thank you for your support.

camolina2 commented 3 months ago

Hi @hvkwak, in which file did you change the params? can you write it here for me please.

Thank you in advance!