1. Creates a flexible model, which is learned from example images (database images).
2. Minimizes error between unknown image and model, determining linear coefficients and affine parameters.

ASM  (Active Shape Model), (Cootes and Taylor): similarly to this approach, it creates a model based out of a linear combination of models, but the initial step is that a user has to select control points on each database image (time consuming, error).  As a correspondence function between each image, the ASM uses a principal component analysis on matrix containing vectors of points of the database images.  Resulting Eigenvectors describe the directions of greatest variations of the control points.  The model is matched by an algorithm that searches a region in a unknown image around the control point to find a better fit, and updates parameters.

Motion Estimation (Hanna and Hingorani): Defines error function which is minimized, to find optimal flow field between images. Sum of squared differences between one image and a warping of the other image, related to the current estimate of the flow field.