search

dunbar12138 / DSNeRF

Code release for DS-NeRF (Depth-supervised Neural Radiance Fields)
https://www.cs.cmu.edu/~dsnerf/
MIT License
749 stars 126 forks source link

Cannot reproduce baseline on DTU #54

Open PruneTruong opened 1 year ago

PruneTruong commented 1 year ago

Hi, thanks for your nice work! I want to compare to it, and therefore I am trying to reproduce the results stated in the paper for NerF (baseline) on DTU, when training on 3, 6 and 9 views. According to the paper, you follow the evaluation protocol of PixelNerf. Therefore, I added in the load_dtu part:

train_idx = [25, 22, 28, 40, 44, 48, 0, 8, 13] exclude_idx = [3, 4, 5, 6, 7, 16, 17, 18, 19, 20, 21, 36, 37, 38, 39] test_idx = [i for i in np.arange(49) if i not in train_idx + exclude_idx] split_indices = {'test': test_idx, 'train': train_idx}

if args.train_sub is not None or args.train_sub > 0: train_idx = train_idx[:args.train_sub]

According to pixelnerf, test split is on 25 views, and the 3 training views correspond to index [25, 22, 28], 6 to [25, 22, 28, 40, 44, 48] and 9 to [25, 22, 28, 40, 44, 48, 0, 8, 13]. Is this what you also did?

For the config, as baseline nerf trained on 9 views, i used the following:

expname = standard_nerf dataset_type = dtu train_sub = 9 render_test = True N_rand = 1024 N_samples = 64 N_importance = 128 use_viewdirs = True raw_noise_std = 1e0 no_ndc = True colmap_depth = False depth_loss = False sigma_loss = False depth_lambda = 0. sigma_lambda = 0. i_testset = 5000 i_video = 0 N_iters = 150000

Yet, when I train on dtu 9 views, I obtained bad results, far from the ones in the paper. For example, on scan40, psnr on the test set is 12. On Scan41, it is 11 and on scan45, 16.

Am i missing something? Thanks for your help.

dunbar12138 commented 1 year ago

Hi, thanks for your interest!

Yes, we follow the evaluation protocol of pixelNeRF and your train/test set up looks good to me.

Did you use the given camera poses in the dataset? Inaccurate poses could lead to bad results.

PruneTruong commented 1 year ago

Could you provide the configs for baseline and DS-Nerf on dtu 3/9 views please? It would be much more convenient. Thanks a lot.

dachao-ai commented 8 months ago

Could you provide the configs for baseline and DS-Nerf on dtu 3/9 views please? It would be much more convenient. Thanks a lot.

Hello, may I ask if your issue with DTU has been resolved? Could you please share the DTU dataset you created? Thank you very much.