Establishing correspondences is a crucial step in generating statistical shape and deformation models. In this abstract we present a technique to compute dense correspondences across a set of images. In contrast to group-wise registration, where images are iteratively registered to an evolving mean image, our method explicitly uses all pair-wise registrations among the set of images by minimizing their group-wise inconsistency using a regularized least-squares algorithm. The regularization controls the adherence to the original registration, which is weighted by the local post-registration similarity. This allows our proposed method to adaptively improve consistency while locally preserving accurate pairwise registrations. We show that the rectified registrations are not only more consistent, but also have lower average deformation error (ADE) when compared to known deformations in simulated data, and lower target registration error (TRE) in clinical data.