NeRO

NeRO copied to clipboard

Thanks for your contribution and the fantastic work! I tried running your code with the default settings according to your instructions on README. However, I failed to get an output as clear as shown in your paper. My results:

While the results from the paper are:

I would like to know why my reconstructed checkerboard was blurred? Thanks for your help!

Thanks for your interest! I'm re-testing the default setting.

In my previous reproducing experiments, I was able to reconstruct it correctly. On step 15k, the board is not fully reconstructed But on step 85k, the board will be reconstructed

In my experience, the initialization of the SDF field indeed affects the final convergence but the default setting should be able to produce the final results because I've fixed the random seed.

Thanks for your reply! I am also re-testing the default settings. Just to clarify, is the board supposed to be fully reconstructed only after stage 2? or stage 1 is sufficient for reconstructing the fine details of the board?

The board is supposed to be fully reconstructed in Stage I.

Hi, I'm able to reproduce the results using the default setting. Maybe you can recheck if there is anything changed in the config files. Meanwhile, you may also reduce this parameter https://github.com/liuyuan-pal/NeRO/blob/59d1b1f5a50e4e022cd53d77b1a27d8144008825/configs/shape/real/vase.yaml#L16 it means how many steps we fix the inv_s. You may reduce it to 0 or 5000, which would help you to reconstruct the board.

The board should be reconstructed within 50k steps.

Thanks for your help! I clone this repo and run the command python run_training.py --cfg configs/shape/real/vase.yaml again. The only modification is that I replace np.bool with np.bool_ to fit numpy version (1.25) I used. However, my results are still blurry this time:

Because I use pytorch 2.0.1, which is different from yours. To check if the problem is caused by python env, I create a new conda env with pytorch 1.11 and rerun the code, but the results are still a bit blurry:

I don't know the exact reason to cause this inconsistent results but you may try to reduce the parameter freeze_inv_s_step to 5000 or 0, which would improve the reconstruction quality of the board.

The reason for this phenomenon is that the appearance (shading) networks are too strong so it is able to fit the appearance on incorrect geometry. While freezing the inv_s makes the geometry converge slower so that appearances are fitted on incompleted geometry.

OK, many thanks for your suggestion. I will try it out.