Am I generating colors correctly?

[Kyle McDonald]

I would like some advice on extracting a voxel representation for generating colors.

I was able to extract very poor vertex color by editing these lines in the G.sample_mixed loop in gen_videos_proj_withseg.py:

sample_result = G.sample_mixed(...)
sigmas[:, head:head+max_batch] = sample_result['sigma']
color_batch = G.torgb(sample_result['rgb'].transpose(1,2)[...,None], ws[0,0,0,:1])
colors[:, head:head+max_batch] = np.transpose(color_batch[...,0], (2, 1, 0))

If I look for the nearest color on the isosurface mesh, it gives me this:

But when I look at the render I see this:

I realize that the render has a final superresolution pass that makes it so clear, but I feel like I might be missing something.

My understanding of the process is something like:

1. G.sample_mixed takes the samples (xyz coordinates in a 3d grid) and the transformed_ray_directions_expanded (which is just 0,0,-1) and w (which is the latent vectors of shape (14,512) from the mapping network output, combining latent and camera pose) and then outputs a few results (sigma, rgb, and a copy of xyz).
2. The rgb is not actually rgb, but it is a 32 dimensional feature vector. So we have to decode it to RGB using the G.torgb network. This is what I find tricky. The network seems designed to process 2D images, but here we only have a bundle of N=10M feature vectors. So I pass it in a 10Mx1 image, and I hope this is ok. Also, torgb expects only a single w from the 14 options. I just picked the first one ws[0,0,0,:1] but I'm not sure if this is correct. Would it be better to run torgb for each w and then average them, or find the median, or something else?
3. Finally, I convert these resulting colors back to voxel space and then use the mesh vertex locations to lookup the closest color.

My questions are:

1. Is it ok to give torgb a 10Mx1 image or is this damaging the performance of the feature-to-color conversion?
2. Is it ok to only use the first ws or should I be using multiple ones somehow? Are each of the ws latents representing a different camera pose, or do they represent something else?

Thanks @SizheAn!

[pengyuhang]

你好，请问下你这个3d模型的颜色是如何生成的，为什么我的3d模型没有颜色

[MustafaHilmiYAVUZHAN]

mix colors according to the 'normal' vector. 2 images are enough to create. full back and front.

[Mohamedhoussein]

Cher MustafaHilmiYAVUZHAN, Je vous remercie de votre email. Je suis heureux de vous aider avec la génération de couleurs. Si vous avez besoin d'une explication plus détaillée, n'hésitez pas à me le faire savoir. Je suis à votre disposition pour vous fournir toutes les informations nécessaires. Cordialement, Housseinelmi Mohamed Le sam. 29 juil. 2023 à 11:53, MustafaHilmiYAVUZHAN < ***@***.***> a écrit :

…

mix colors according to the 'normal' vector. 2 images are enough to create. full back and front. — Reply to this email directly, view it on GitHub <#27 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/A7FJOG7P3Y4GXVIXHE5ZLUTXSUWVZANCNFSM6AAAAAA2XUEDVQ> . You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>

[Miles]

Hi, how did you get the RGB texture ?

[Kyle McDonald]

mix colors according to the 'normal' vector. 2 images are enough to create. full back and front.

@MustafaHilmiYAVUZHAN thanks for your input. Do you have any reference code for this? When you say normal vector, do you mean the w vector? Should I run multiple w vectors through torgb and then take an average or something?

[MustafaHilmiYAVUZHAN]

view uv
@kylemcdonald

[Xinzhou Wang]

Hi! Very interesting attempt! Could you please share the code of "look for the nearest color on the isosurface mesh"? Thanks a lot!

[Xinzhou Wang]

Followed by this code, I got the mesh, but the color seems not right.
Notice that the colors is in range[-3.1344, 3.2253] so i clamp it to [-1, 1]. Then i turn it into [0, 1]