Standard convolutions are prevalent in image processing and deep learning,
but their fixed kernel design limits adaptability. Several deformation
strategies of the reference kernel grid have been proposed. Yet, they lack a
unified theoretical framework. By returning to a metric perspec