[Enchancement] A bug in original Gsplat. Maybe it also happen in this implementation

[ichsan2895]

Today, I have read discord that gsplat & nerfstudio has a big bugs. After it fixed, it being said that the quality is increasing.

The bug is about The calculation of cov2d's grad is incorrect and the unit test doesn't report errors because the test cases are too weak: https://github.com/nerfstudio-project/gsplat/pull/136

It claims the fixing will be increasing the quality of reconstruction. Here is the discord link: https://discord.com/channels/1025504971606724712/1167872346707738664/1212867101585571850

[ichsan2895]

Another news: Seems it fixing a bug but get another bug. The result is worse than before. I will wait for it until all bugs is dissapears

[bchretien]

@ichsan2895 indeed, if you want to test I've got a quick (untested) patch:

0001-Fix-from-gsplat.patch

I've got a training running so I will now in a while if it's ok. It's still pending further investigations from the gsplat devs as there may be more on the way as you said.

[bchretien]

@ichsan2895 there it is:

• https://github.com/nerfstudio-project/gsplat/pull/139
• https://github.com/nerfstudio-project/gsplat/pull/140

so patch above to be updated.

[ichsan2895]

Thanks for that information :) Hopefully if that bugs happen in this implementation, it will be fixed.