[QUESTION] Is it possible to construct a COLMAP dataset from existing point cloud and known camera positions?

[Dmitry-Filippov-Rival]

I am trying to create a Gaussian Splatting representation of my render projects so that I am able to navigate them in 6DoF.

The way I envision the workflow is that I take my existing blender project, render out a few hundred scenes from it with known camera positions, then create a point cloud from the project's mesh and use that to seed the gaussian splatting learning algorithm. Since both point cloud and rendered camera scenes come from the same project, I know the relative positions of everything perfectly and using SfM to reconstruct the scene seems wasteful.

All the Gaussian Splatting training algorithms I can find only take COLMAP or similar datasets as input, so I wonder if it is possible to synthetically construct the dataset without going through the SfM reconstruction and what would it take to do that?