Figure 5: Left to right: Patch positions after zero, one, two and final iteration with line-search (upper row) and without
line-search (lower row).
the weight matrix W = I, this corresponds to the resid-
ual F(x) being orthogonal to the plane spanned by the
columns of J(x), as illustrated in Figure 4. If W # I,
the vectors will instead be conjugate with respect to W.
Near the solution, the angle 9 between J ; p; and F'; will
approach 7/2 and the quotient |LJ x pk||/||LF || = cos 8
will approach zero. Thus, a termination criteria on 8 may
be formulated as
ILI pi|| € &|[LF||
for a constant € = cos ! (1/2 — fiiit).
This termination criteria will work for almost all problems
with a non-zero optimal residual. However, for very small
optimal residuals, e.g. for synthetic, perfect data, the ter-
mination criteria may have to be adjusted to
|LJzpzl| < + e)|[LFr||-
to avoid unnecessary iterations due to round-off errors.
Of course a heuristic problem-dependent termination cri-
teria may be used as well.
5 SIMULATIONS
5.1 Example iteration
The Gauss-Newton method with and without line-search
was applied to a least squares template matching problem
where the template was a 11 x 11 pixels white square with
a4 pixel black border. The image consisted of tiled squares
19 pixels wide with a separation of 12 pixels. The template
was binary (black=0, white=255) while the image was low-
pass filtered in order to create a more difficult matching
problem. The transformation had 8 parameters, allowing
affine deformation and radiometric changes both in scale
and shift. The picture re-sampling function was bilinear.
Implementation details are found in (Borlin, 2002).
A typical example of the algorithmic differences is shown
in Figure 5. In the first iteration both algorithms take a full
step (first and second column). In the second iteration, the
undamped algorithm takes the full step (lower third col-
umn) and starts to diverge. After 20 more iterations, it os-
cillates about the position shown in the lower right column.
In contrast, the damped algorithm rejects the step lengths
a = 1 and a = 0.5 since the residual norm would increase,
but accepts a = 0.25 (top third column), and converges in
six iterations toward the optimal solution (upper right col-
umn).
5.2 Pull-in range
The effect on the pull-in range was investigated by ap-
plying the two algorithms with progressively more distant
starting approximations and recording their success or fail-
ure.
The optimal patch parameters were a11 = b11 = 57.95,
aı2 = b91 = 1.56, 5] — b19 = 0, Tg — 34.33, rt = 0.1.
The starting approximation of the patch position a11,511
was varied between 45 and 71. The starting approxima-
tions of the other geometric parameters were a1» — 051 —
1.5 and a91 = bız = 0. The starting values for the radio-
metric parameters were approximated from the picture at
the initial patch position.
The pull-in ranges and side minima convergence for the
algorithms are illustrated in figures 6 and 7, respectively.
The undamped algorithm converged toward the correct so-
lution for 223 starting positions while the damped algo-
rithm succeeded for 326 starting positions. The undamped
algorithm converged toward a side minima at 5 occasions.
The corresponding number for the damped algorithm was
48.
6 SUMMARY
These simulations illustrate the usefulness of a line-search
strategy in solving an Adaptive Least Squares Matching
problem. In the example iteration, the damped algorithm
avoids divergence at the cost of two extra residual calcu-
lations in the second iteration, a cost which is more than
balanced by the fourteen extra iterations the undamped al-
gorithm performs before giving up. For the same problem,
the pull-in range for the damped algorithm is higher than
for the undamped algorithm. However, the repetitive pat-
terns in the example image also increased the convergence
towards identical side-mimina. This is likely to be less of
a problem in less non-repetitive images.
—10—
Figure 6:
undampe:
for which
converge
Finally, tl
potential
two resez
squares o
plored in
REFERI
Bórlin, N
Gauss-Ne
Departme
Dennis Ji
Methods
Equation:
Eriksson,
Nonlinea:
ment of C
Eriksson,
P-A., 19
forward 1
Optimiza
Gruen, A
tal measi
Close Ra
tles, Caitl