A new approach for the automatic estimation of dense 3D deformation fields is proposed. In the first step, template propagation (an advanced block matching strategy) produces not only a large set of point correspondences, but also a quantitative measure of the “registration quality” for each point pair. Subsequently, the deformation field is obtained by a method based on Wendland radial base functions. This method has been adapted to incorporate “registration quality” into regularization, where Morozov’s discrepancy principle has been applied to give intuitive meaning to the regularization parameter. The main advantage of the presented algorithm is the ability to perform an elastic registration in the presence of large deformations with minimum user interaction. Applying the method, complicated respiratory motion patterns in 3D MR images of the thorax have been successfully determined. The complete procedure takes less than one hour on a standard PC for MR image pairs (256×256×75 voxels) showing a 40 mm displacement of the diaphragm.