Let be an affine camera motion obtained by factoring the measurement matrix. Then is related to the ``true'' motion described with respect to the Euclidean coordinate frame by (9). Our goal here is to obtain the affine transformation (up to an arbitrary rotation) in (9) with which affine descriptions of motion and shape are corrected to Euclidean ones .
Let two 3-vectors and denote two rows of . Then, for each affine camera model introduced in §2.2, the affine transformation is determined as follows;
Let and be two elements of . Then, eliminating from (19), we obtain
which gives 2F linear homogeneous equations in terms of 6 elements of the 33 symmetric matrix . If three or more views are available, we can obtain a least-squares solution determined up to a scale as a solution of the ordinary eigenvalue problem. Once is determined, is recovered through, for instance, Cholesky decomposition.
and the affine transformation is determined as a least-squares solution of the following equations:
and is recovered by solving, in the least-squares sense,