The proper way of getting C2W matrix from self.world_view_transform?

[justin4ai]

Hi,

thanks for your amazing code work, we are able to access to W2C matrix(= self.world_view_transform) here.

However, the code for the function getWorld2View2 is very tricky; The function that I'm confused with the way of getting the proper C2W matrix from self.world_view_transform.

Do I need to 1. transpose 0th and 1st dimension against self.world_view_transform and 2. apply the inverse function (just following the exact reverse step)?

Thanks,
Junyeong Ahn

[sonbosung]

It seems that the transpose operation is just for following glm convention(column-major order). The diff-gaussian-rasterization submodule uses glm library.

[justin4ai]

Thanks for your help. Then self.world_view_transform.T must stand for the right world to camera matrix (based on row-major Python). Applying inverse function after the transpose may lead to C2W matrix.

[sonbosung]

You're correct. Also, you can note a potentially confusing expression in the getWorld2View2 function, where it transposes the given rotation matrix R. When reading the dataset in the readColmapCameras function in dataset_readers.py, it assigns the transposed rotation matrix to R.