Precision improvement

The estimated parameters have the same precision as the shapes, i.e., approximately one pixel
We try to minimize
$$\arg\min_{s, \theta, \mathbf{t}} \sum \Vert s R_\theta~\mathbf{x}_i + \mathbf{t} - \mathbf{y}_j \Vert^2$$
Linear problem with unknowns $s\cos\theta$, $s\sin\theta$, $t_x$ and$t_y$
We restrict the summation to correspondences compatible with the dominant motion (its electorate) $\rightarrow$ no influence from secondary motions on precision

Summer School on Mathematical Problems in Image Processing, Trieste, Italy, September 4th-18th 2000