This paper presents a new method for warping of diffusion tensor fields. The proper reorientation of a tensor field requires knowledge of the underlying fiber orientation, which is not known a priori. Accordingly, a probabilistic representation of the fiber direction is adopted and used in a Procrustean estimation of the rotational component of the warping field. The estimated rotational component, along with the warping field, are then used to warp and properly reorient a tensor field. Results from diffusion tensor MR images demonstrate a dramatic improvement of signal to noise ratio after statistical averaging of properly registered tensor fields across 9 individuals.