How to train a Gaussian model using a background-removed dataset?

[KuoFengYuan]

Describe
I use the GS model to train the transparency dataset, but the trained background has many clouds.
Screenshots

[brentyi]

Not my area of expertise, but nerfstudio-project/gsplat#70 seems relevant!

[KuoFengYuan]

thank you for your reply！
I will researching it

[KevinXu02]

A fast but inelegant method is to directly inject white bg colour into GS model, I've test it with LEGO dataset in #2704. And then you can transit the rgb image with white bg to a rgba one. I think the main problem is that when training GS model the bg colour is ramdomly picked, we may need a alpha for the bg and add it into the BP and FP I think? nerfstudio-project/gsplat#70 seems only apply alpha output to gaussians, but I think it will immediately work by adding alpha for the bg and setting it to 0.

[KuoFengYuan]

Yes, I currently use a white background to run training