ajhamdi / ges-splatting Goto Github PK

Hi guys,

thanks a lot for your efforts. I notice that in eq.(1) of the arxiv paper, beta is a learnable parameter, but in the current implementation, it is still fixed to 2:

Seems like even when we are using non-Gaussian kernels (i.e. render_laplacian), we are still drawing Gaussians with covariance matrices linearly scaled by Phi(beta), as described in eq.(4).

Is it true that the rasterizer implementation isn't updated? Or maybe I miss something and misunderstand it.

Hello, Appreciate for nice work.

I try to point cloud that extracted DROID SLAM as input point cloud but Camera pose is extracted by ColMAP

but it is not work.
under 500 iter, it doens't add new gaussian, and over and over iteration, just fade out.
So I visualized the image(rendered), and this image was only black.

Why it is Happend ?

In render.py, why do you use original render function instead of render_laplacian? Although you save ply.file as get._scaling() * get.shape(). but it is different from render_laplacian?

from scene.gaussian_model import GaussianModel
...\ges-splatting\scene_init_.py", line 20
from future import annotations
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
SyntaxError: from future imports must occur at the beginning of the file

Currently when running either GES or GS i receive this issue

Thank you for this awesome work.
I am interested in reproducing figure 9 where you constrain ges to about the same number of splats as the reference implementation. How to constraint the training to reproduce the results?

Thx and best,
Janusch

Hello,

It's a great work, thank you.

I am getting below error, not sure why.

mu1 = F.conv2d(img1, window, padding=window_size // 2, groups=channel)
RuntimeError: Expected 4-dimensional input for 4-dimensional weight [3, 1, 11, 11], but got 3-dimensional input of size [3, 545, 980] instead

Can you please help me to resolve this.

Thank you
Gopi