DTU evaluation accuracy? about mvsnerf HOT 7 OPEN

from mvsnerf.

Hi, have you re-produced the results on DTU dataset? I used the same setting of the author, but I got much lower PSNR than author in paper.

from mvsnerf.

I have the ame question. Maybe authors provide a sub-optimal model for us.
By the way, I re-produced the results on DTU and obtain a higher PSNR than given checkpoint.
psnr 26.673; ssim: 0.931; lpips: 0.172

from mvsnerf.

I get the same numbers as hdjang.
I wonder why the PSNR values without fine-tuning in the paper correspond to results where always the 3 nearest views are used as input for each validation image whereas the description in the paper tells that always 3 fixed views are used

For each testing scene, we select 20
nearby views; we then select 3 center views as input, 13
as additional input for per-scene fine-tuning, and take the
remaining 4 as testing views.

With 3 fixed views, the PSNR values I obtain for the given checkpoint by using the renderer.ipynb are significantly lower (21.05 for DTU).
Could the authors clarify that?

from mvsnerf.

@zhangchuanyi96 Hello, I 'm trying to reproduce the results following the command in ReadMe:
'''
python train_mvs_nerf_pl.py --with_depth --imgScale_test 1.0 --expname mvs-nerf --num_epochs 6 --N_samples 128 --use_viewdirs --batch_size 1 --dataset_name dtu --datadir ../data/mvs_training/dtu
'''
But the results is much lower than the results in paper.
Could you please tell me the command or the hyperparameters you use when re-producing the DTU results? Thank you.

from mvsnerf.

@zhangchuanyi96 Hello, I 'm trying to reproduce the results following the command in ReadMe: ''' python train_mvs_nerf_pl.py --with_depth --imgScale_test 1.0 --expname mvs-nerf --num_epochs 6 --N_samples 128 --use_viewdirs --batch_size 1 --dataset_name dtu --datadir ../data/mvs_training/dtu ''' But the results is much lower than the results in paper. Could you please tell me the command or the hyperparameters you use when re-producing the DTU results? Thank you.

It has been months, so sadly I can't remember the exact reproducing process.
I can only vaguely remember that I probably didn't change the given hyperparameters. Your problems may be related to machines.

from mvsnerf.

I think the setting of the results given by the authors in the paper is to select the 3 nearest views. Fine-tuning is performed on 16 training views. what do y'all think?

from mvsnerf.