ABSTRACT
We assume that low-level processes can provide an adequate initial estimate of scale and location of the input face, so we are generally interested in the question of finding a warp 0(x) which is near an initial guess. Nonetheless, to find the correct 0(x) using a local optimization method like gradient descent with respect to the functional E, corresponding intensity peaks and valleys in the model and data images must overlap, so we are forced to explicitly define 0(x) as the composition of a global (affine) transformation #(x) used to adjust the position, scale, and rotation of the face and a local transformation £(x) used to adjust the precise boundaries and internal features of the face. By finding the best affine transformation first, we considerably raise the probability that the local warping will correctly perform fine-scale tuning. Similar ideas are developed in the thesis by Matevic (97).
