Question on Rotation Methods in the Paper

[jk4011]

Thank you for your excellent work.
I really enjoyed reading your paper.

I have a question regarding the rotations mentioned in your paper.

Specifically, your paper introduces two types of rotations: the MLP-based rotation in Eq.4 and the SVD-based rotation in Eq.9.

I’m curious why you introduce the SVD rotation instead of relying solely on the MLP rotation. Wouldn’t the MLP rotation alone suffice to maintain the ARAP (as-rigid-as-possible) principle?

Could you please clarify if there’s something I might be misunderstanding?

Thank you in advance for your time and help.

[yihua7]

Hi, sorry for the late reply.

• The MLP-predicted rotations govern the motion and shape of the Gaussians, capturing the relationship between the control points and the Gaussians.

• In ARAP, the trajectories of neighboring control points are considered (control points to control points). A control point and its neighbor form a directional graph, where the rigid rotation can be estimated using SVD. Instead of focusing on the relationship between control points and Gaussians, ARAP penalizes the relative non-rigid deformations occurring between the control points.