Paper says positions and covariances aren't learnable, but they seem to be in the code?

[zeki31]

Hi, thank you so much for your great work — it's really impressive!

Possibly I may have found a small inconsistency between the paper and the official implementation regarding 2DGS optimization with Charts.
In the paper (Section 4.4), it says:

All other parameters of the Gaussians, such as positions and covariances, are not learnable and computed on the fly depending on the position of the vertices.

However, in the provided code (e.g. https://github.com/Anttwo/MAtCha/blob/main/2d-gaussian-splatting/scene/gaussian_model.py#L179), it seems that both positions and covariances are still learnable and updated through optimization.

Is there something I might be misunderstanding? If the paper's description is correct, I would really appreciate it if you could point me to the relevant part of the implementation.

Thank you for your time and consideration!

[xiyichen]

Same question. I tried to remove them from optimizable parameters and there seems to be performance degradation on my data. The PSNR dropped by ~2.

[DaLi-Jack]

Same question.