Next: Depth estimation using
Up: Factorization under perspective
Previous: Factorization under perspective
In the case of perspective cameras, we rewrite
(1) for the all views and points as
If all the depth parameters are known, we can construct
a measurement matrix and decompose it
into camera motion and object shape using
SVD in a similar manner as affine cases. This factorization has the
following ambiguity; Let be the projective depths which
give a measurement matrix that can be decomposed into motion and shape
as in (10). Then the following
equation holds for any 44 non-singular matrix
and any non-zero scalars and
:
where
From these two equations, we can observe two things:
- Ambiguity of depths:
- If one family of projective depths
{} yields a measurement matrix that can be
decomposed into motion and shape, another family
{} given by (11)
also yields a decomposable measurement matrix. Therefore,
without loss of generality, we can choose one view, say
i=1, and one object point, say k=1, and fix the depths
associated with this view and this point to unity:
Consequently, the number of independent depth parameters is
.
- Ambiguity of reconstruction:
- Assuming that we obtain two
solutions and
by factoring two
measurement matrices constructed from two depth families
{} and {},these two
solutions are then related by using an unknown non-singular
matrix and unknown scalars and as
Equation (13) means that we can recover
motion and shape only up to an unknown projective transformation and can
arbitrarily choose a 3D projective coordinate frame in terms of which
structure of the motion and shape are
described[2,4].
We call this kind of recovery projective reconstruction.
Bob Fisher
Wed Apr 21 20:23:11 BST 1999