In this chapter we will study so called matrix factorizations or matrix decompositions.
These are different, and useful, ways of writing a matrix as a product of matrices.
These factorizations are extremely useful for understanding matrices and for making computations
with matrices.
We start with an example from the application of computer vision.
Example 11.1:Structure from Motion using Matrix Factorization
An approximative model for the image projection of a 3D point with coordinates $(x,y,z)$ to an image point with coordinates $(u,v)$
is the so called affine camera model
Assume that $n$ points with column vectors $\vc{x}_1, \ldots, \vc{x}_n$ are all seen in $m$ images
and denote by $\vc{u}_{i,j}$ the column vector $\vc{u}$ for point $j$ seen in image $i$.
Let the camera matrix for the $m$ images be $\mx{P}_1, \ldots, \mx{P}_m$.
The $mn$ constraints can now be written as one big matrix equation
The matrix to the left $\mx{U}$ contain the measured (known) image coordinates that have been detected in the images.
To the right we have the product of two matrices $ \mx{M}$ and $\mx{X}$ that contain the parameters that we would like to estimate.
If we only could factorize or decompose $\mx{U}$ as $ \mx{M} \mx{X}$, this would solve our problem.
Popup-help:
An $n$-dimensional column vector, $\vc{v}$, is represented with respect to a basis and consists of a column of $n$ scalar values.
The vector elements are sometimes denoted $v_1$, $v_2$,..., $v_n$. For two- and three-dimensional vectors, we sometimes
also use $v_x$, $v_y$, and $v_z$.
The notation is
where $\vc{u} = u_x \vc{e}_1$, $\vc{v} = v_x \vc{e}_1 + v_y \vc{e}_2$,
and $\vc{w} = w_x \vc{e}_1 + w_y \vc{e}_2 + w_z \vc{e}_3$.
Note that $\vc{e}_i$ are the basis vectors.
In our text, we also use the shorthand notation $\vc{w} = \bigl(w_1,w_2,w_3\bigr)$, which means the same as above
(notice the commas between the vector elements).
A row vector does not have any commas between, though.
Popup-help:
Projection is a mapping from one set onto a smaller set (i.e., a subset) of that set. For example, if a 3D point
is projected to the closest point on a plane, then starting set was all 3D points which was reduced to
the points in the plane using projection.