Concerns about the initial number of 3D Gaussians

[Time-Lord12th]

Hi, thanks for the great work. The paper mentions D=128, and "After obtaining the multi-view depth predictions, we directly unproject them to 3D point clouds using the camera parameters." So is the initial number of 3D Gaussians 128*K? Isn't this number too small? Approximately how many 3D Gaussians are there after the training is completed?

[donydchen]

Hi @Time-Lord12th, thanks for your interest. The number of 3D Gaussians is $H \times W \times N$, where $H, W$ are the image resolutions, and $N$ is the number of context/input views. The number of 3D Gaussians is fixed; we only update the properties of the Gaussians but not the quantity. And we find that such a number is enough for our experimental datasets.

Edited: Correct the number of Gaussians. Thanks for the below reminder from @Time-Lord12th.

[Time-Lord12th]

Thanks for your fast and detailed reply!

[Time-Lord12th]

@donydchen Sorry to bother you again. The number of 3D Gaussians should be HxWxN? As stated in the paper：

Could I possibly ask another question? For ablations, when removing cost volume, how does MVSplat predict 3D Gaussian parameters, does it predict probabilistic depth like pixelSplat? Thanks!

[donydchen]

Hi @Time-Lord12th. Yes, you are correct. Sorry for the mistake I made above. It is pixel-aligned and should be just $H \times W \times N$. The depth dimension will be reduced to 1 via weighted average using the weights obtained via softmax, implementation details are here.

---

When removing cost volume, we simply rely on the feature to regress the Gaussian centers, without using any probabilistic sampling like pixelSplat. Related implementation details are here.

---

Feel free to ask any questions regarding our MVSplat; we would be more than happy to answer them. Thanks.