poetrywanderer / cf-nerf Goto Github PK

Hello,

Thank you again for posting the code.

The following lines of code in run_nerf_uncertainty_NF.py refer to the key 'network_fine' in the dictionary render_kwargs_train, even though this key is never created resulting in a KeyError.

        if i%args.i_weights==0:
            path = os.path.join(args.basedir, args.dataname, args.type_flows, args.expname, '{:06d}_{:02d}.tar'.format(i, args.index_ensembles))
            if args.N_importance > 0:
                torch.save({
                    'global_step': global_step,
                    'network_fn_state_dict': render_kwargs_train['network_fn'].state_dict(),
                    'network_fine_state_dict': render_kwargs_train['network_fine'].state_dict(),
                    'optimizer_state_dict': optimizer.state_dict(),
                }, path)
            else:
                torch.save({
                    'global_step': global_step,
                    'network_fn_state_dict': render_kwargs_train['network_fn'].state_dict(),
                    'optimizer_state_dict': optimizer.state_dict(),
                }, path)
            print('Saved checkpoints at', path)

Note that this key is explicitly created in the run_nerf.py file here: https://github.com/yenchenlin/nerf-pytorch/blob/master/run_nerf.py from the original nerf-pytorch repository as shown below:

        'network_query_fn' : network_query_fn,
        'perturb' : args.perturb,
        'N_importance' : args.N_importance,
        'network_fine' : model_fine,
        'N_samples' : args.N_samples,
        'network_fn' : model,
        'use_viewdirs' : args.use_viewdirs,
        'white_bkgd' : args.white_bkgd,
        'raw_noise_std' : args.raw_noise_std,
    }```

Hello,

Thanks for sharing the work. We are trying to follow your work in our paper.
However, I am a little puzzled by the calculation here $\sigma_{rgb} * (0.8/n)^{-1/7}$

The expression is corresponding to the following code.

Could you elaborate more?
Thank you.

Best,

I was wondering how you performed the 20% train - 80% test split in LLFF dataset? I cannot find a direct code line that does this for LLFF dataset in the repo. I can see that for the LF dataset about 4 views are chosen as training, but I was wondering how you do it for LLFF? If possible could you share the code for this split (training indices)?
I noticed that S-NeRF has this 20-80% split for train/test, but you say in the paper that for test set you use the the last adjacent views.

we use a sparse number of scene views (∼4) for training and the last adjacent views for testing.

What is meant by 'the last adjacent views'? Is it the remaining 80% of the data or something else? @poetrywanderer

Would you be able to upload the pre-trained models? There is a section in the main page for it but it is empty. Thanks!

Hi,

Thanks for releasing the code! I believe the highlighted part of the Anaconda environment file should read "yml" not "yaml."

Hi,
I'm trying to reproduce the results from table 2 without success. The code seems to no longer have a test function, which should be executed without passing --is_train as an argument. In fact, the test script test_NF.sh does not perform any computation and exits immediately. Additionally, the model weights mentioned in the Pre-trained Models section in the README.md seem to be point clouds rather than the weights of the NeRF networks.

How can I reproduce the results from table 2?

This is a very good job. Thank you for sharing. But I have two questions, one is how to visualize image results like your paper.
An other question is when I run the test_NF.sh, it seems that there is no output

Hi,

In the code, in line the errors are not normalized (commented out). I was wondering why this is and if the metrics reported in the paper were reported without normalization.

Thank you!