Training details for NeuMan dataset about ml-neuman HOT 3 OPEN

In the background training stage, we used 4096 rays+128 samples+0.5M iterations. In practice, you can lower the number of samples to 64. Depending on the VRAM, you can also lower the number of rays. The configuration is flexible, just pay attention to the validation visualizations to make sure the depth and RGB renderings look good.

from ml-neuman.

Got it, thanks for the quick reply!

from ml-neuman.

Hi Wei, it seems that training the combined model for 50k steps takes much longer time than 5 days (3-5 seconds per iteration on 3090).

To reproduce the results in the paper, do we need to train the BG model for 50k steps and the combined model for 50k steps?

from ml-neuman.