Scaling transformation for Gaussians by control point deformation

[Friedrich-M]

Thanks for your great work! I wonder why the control point deformation network will output the scaling transformation for Gaussian points which is not indicated in the main paper.

################## in the deformation function ##################
scale = (node_scale[nn_idx] * nn_weight[..., None]).sum(dim=1) * motion_mask
return_dict = {'d_xyz': translate, 'd_rotation': rotation, 'd_scaling': scale}

################## in the rendering function ##################
scales = pc.get_scaling + d_scaling 
rotations = pc.get_rotation_bias(d_rotation)

Since you adopt LBS to guide the transformation for Gaussian points, I am curious why control points will affect the scaling attribute of Gaussians.

@yihua7 Thanks in advance!

[zhenyuan1234]

I wonder about that, too. Are you understanding this?