kaolin-wisp
kaolin-wisp copied to clipboard
NVIDIAGameWorks
How to reproduce the results in the paper (instant-ngp)?
I train the ngp-nerf with conmand
python3 app/main.py --config configs/ngp_nerf.yaml --multiview-dataset-format standard --mip 0 --dataset-path data/lego
I only changed epochs to 100
The PSNR of result is only 24.52 in the lego dataset.
How can i reproduce the results in the paper (instant-ngp)
Same question here! I trained on the lego scene with the provided configuration file configs/ngp_nerf_interactive.yaml, but the results look a bit blurry and not as good as what I got from instant-ngp.
I still cannot reproduce results using the latent version of the wisp. Does anyone know the correct hyper parameters?
@orperel has there been any update on how to reproduce the results from the original instant-NGP paper? The current version of wisp doesn't match the performance.
Hi! We worked around this issue, but we need to further clean this MR before we can merge it. Are you blocked for ICCV? If so, we can post the dirty MR until this effort is done.
Otherwise, the metrics in the NeRF table should be up to date. (after we merge PSNR on lego should go above > 35.0)
@orperel Yes, please post/share the required changes needed. I do indeed need it for ICCV.
@tovacinni Has there been any update on this? I would really appreciate if the results from the original papers can be reproduced. Otherwise, it is hard to use this framework to do experiments related to publications. Thanx for your hard work.
Hi, this has been still blocked on some other MRs that need to be cleared. Basically to explain what's missing if you want to implement them yourself:
- Adaptive ray batching from INGP
- Some hyperparameter tuning
- The use of huber loss
- Some other small tweaks that aren't in the INGP paper but are in the code
All of these things can improve the results.
@tovacinni If you can provide specific values for these hyper parameters, that will be great. Thanx.