For the appearance quality calculation, we should first align the 3D coordinates of the centers of decoded gaussians with the gt point cloud, then the rendered images can be compared fairly.

My question is if my understanding of the evaluation process is right, how do the aligning process? and how the gt point cloud is attained?

btw, you said you use the Filter-Reg instead of the ICP to do the aligning process, could you share which repo you use to do the filter-reg?