About the degenerate solution

[THHH0Sunshine]

In the paper, the 2D Gaussian is lower-bounded by a fixed-size Gaussian in screen space:

This results in very thin Gaussians having a dot at its center instead of being low-pass filtered like 3DGS:

Thin 2D Gaussian without the degenerate solution:

Thin 2D Gaussian with the degenerate solution:

Is this intended? I think that the solution may not solve the issue of gradient instability, because the needle shaped Gaussians still have large gradients, which causes over densification in certain areas:

The over densification issue is also mentioned in CityGaussianV2. The issue does not occur in 3DGS.

[hbb1]

Hi, I didn't follow and I didn't encounter such issue before. The low pass filter mainly does magnification to the reconstruction filter. What do you mean gradient instability and how is it related to densification?

[THHH0Sunshine]

The issue is that there are many degenerated Gaussians with large gradients, and Gaussians with large gradients are cloned or split during densification, leading to number explosion of Gaussians. This does not happen on 3DGS, so I think it may be related to the filtering strategy difference between 3DGS and 2DGS.

[hbb1]

I guess this is more related to the densification strategy since we adopted a different method (but we intend to mimic that of 3DGS) for increasing points. The filtering strategy is also possible, but inspecting densification strategy is a good starting point for you. I believe that you would benefit from discussing with the authors of CityGaussianV2.

[THHH0Sunshine]

Thank you for your answer.