STNs are meant to spit out the parameters of the affine transformation that’ll be applied to the input feature map. It is usually applied on 2D data where the affine transformations have clear definition. Is there any work that extended this to 3D data like meshes and graphs ?