Number of gaussian and shape

[kelo020304]

I am really confused :
1)why number of gaussian sphere seems little and achieve such dense voxel representation
2) why all gaussian seems a sphere rather than Ellipsoid, I think this do not take advantage of gaussian representation.

[JunShao0104]

Me too, did you figure out what's going on?

[seamie6]

Which part of the code did you use to generate this?
Also, given this, it seems redundant to optimise scale and rotation if the Gaussians remain spherical, due to spheres being invariant under rotation about their center point.

[kelo020304]

Me too, did you figure out what's going on?

NO，what do you think about it

[kelo020304]

Which part of the code did you use to generate this? Also, given this, it seems redundant to optimise scale and rotation if the Gaussians remain spherical, due to spheres being invariant under rotation about their center point.

I just use visualize.py to get this pic with 25600 gaussian, and I tried to close the restrict_xyz, but the model do not converge on 4090 for 3days

[SOTArt]

Which part of the code did you use to generate this? Also, given this, it seems redundant to optimise scale and rotation if the Gaussians remain spherical, due to spheres being invariant under rotation about their center point.

I just use visualize.py to get this pic with 25600 gaussian, and I tried to close the restrict_xyz, but the model do not converge on 4090 for 3days
hey, I set the batch_size as 1 in ./config/base/surroundocc.py, and I used 4 pieces of 4090, but the memory was not enough. Have you met this problem and how did you to solve it?

[kelo020304]

this

I just use one 4090 and I can train 25600 version

[SOTArt]

this

I just use one 4090 and I can train 25600 version

oh thanks, I will try again

[seamie6]

Hello, I can see the reason why the number of Gaussians has reduced greatly
Inside vis.py we have save_gaussian() which visualise the Gaussians. I am running an altered model with 51200 Gaussians but the fundamentals stay the same. We see the line

    mask = (pred != empty_label) & (opas > 0.75)
    means = means[mask]

This removes any Gaussian that has an opacity of less than 0.75. In my rendition, this decreases the Gaussian number from 51200 down to 2k to 5k. Huge decrease in the number. I got this via a print statement before and after this mask was applied

print('Gaussian number:', means.shape)

I assume this is done for quicker visualisation. You can experiment with this mask so we visualise all Gaussians, for which we would change a snippet to opas>0