real scene test, get poor result

[neilgogogo]

hi, thank you very much for your work.

I have one question, when I try to test my real scene data. When it comes to evaluate, the result is like this,

but when it comes to novel view, the result is the following image,

What the reason behind this result do you think?

Many thanks.

[ashawkey]

@neilgogogo Hi, have you tried the GUI, so you can visualize which regions are blurred?

[neilgogogo]

hi, thank you for your reply.
I try to run it on windows, but get the following error when compling the cuda

[ashawkey]

Could you give more details on the enviroment? e.g., CUDA version, VS version, ..
also check this thread.

[neilgogogo]

I tried to downgrade the environment to cuda11.3.
And it comes the following compiling errors.

windows10
Cuda compilation tools, release 11.3, V11.3.109
Python 3.8.5
torch1.11.0+cu113
VS2019
RTX3060

[ashawkey]

The error is strange... Could you post the full error log, or at least the first error? Also, can you install tiny-cuda-nn?

[neilgogogo]

The error is strange... Could you post the full error log, or at least the first error? Also, can you install tiny-cuda-nn?

sorry, it is really a long error, and the start is limited by the powershell. Perhaps it is still due to the environment issue when compling the hashencoder.

[ashawkey]

@neilgogogo Hi, In fact I found some related issues:

• windows10 build error  NVlabs/instant-ngp#178
• MSVC 14.29.30037 with CUDA 11.2 occurs error expected a "(" actions/runner-images#3485

You may have a try at these workarounds:

• change /std=c++14 to /std=c++17 in all backend.py
• downgrade VS toolset by setting -T version=14.25 for cmake
• upgrade both VS and CUDA to the latest version

[neilgogogo]

hi, thank you for your advice. When I change C++14 to C++17, it works;)

[neilgogogo]

hi, when I visualize the training, the view is like this. I think the reason might be the camera coordinate? But I am not sure.

Is it possible to visualize the training camera poses and bounding box in dearpygui, so it might be more easier to debug.

[ashawkey]

It's not implemented now, but telling from the second figure, the camera poses are not properly transformed. They should be centered around the unit sphere, an example:

You could try running colmap2nerf.py on the original images, or if you already have the poses, you could refer to these parts to properly adjust them.

[neilgogogo]

hi, using the tensoRF, I reconstructed the target. But the reconstructed resolution is not good after 1000epochs*64 steps. I find there are resolutions 0 and 1 in the main function. Do you think increasing it will increase the resolution? Or how to increase the image resolution do you think， such as increasing the sigma_rank, color rank.

[ashawkey]

64 steps sound too few... I thought the default is 512 steps?

[neilgogogo]

thank you for your advice, I will try;)

cheers