Assess whether automatic regularization inference extends to dense 3D deformation fields

Determine whether automatically inferring the regularization strength in probabilistic registration extends reliably from low-dimensional deformation models to dense three-dimensional deformation fields with millions of effective degrees of freedom.

Background

The proposed sampler can in principle sample the regularization strength, thereby eliminating the method's remaining user-tuned hyperparameter. Prior work has successfully performed automatic regularization inference, but only for low-dimensional deformation models with relatively few parameters. The paper raises uncertainty about whether that success carries over to dense 3D deformation fields, where very low regularization strengths may permit millions of effective degrees of freedom and substantially degrade registration accuracy.