about training precedure cost?

[chensjtu]

Thanks for your great work! I wonder about the time cost, iteration numbers, and device when training on DTU dataset. Would you please tell me this important information?

[Xharlie]

hi, it is similar to MVSNeRF, which takes 250000 iterations, almost 8 epochs. However, after 3-4 epochs, the growth is starting to be slow. it takes me about one day, but greatly depends on your cpu/gpu. Please check the scripts in dev_scripts/ete/ folder for details

[yjcaimeow]

Hi, I understand that the total time consumption should be 1 day for init plus 2 minutes for each scene?

[chensjtu]

Thanks for your quick reply! So it takes about one day to train the init MVSNET as well as MLPs in point-nerf? After the init training, these networks can be directly used in inference and maybe later per-scene optimization?

[Xharlie]

yes, however, without initial training, you can also directly do per-scene optimization from random value, but the convergence is slower.