Rob Reeves
Mean Error in Disparity for Converged Match Windows Mean Error in Disparity for Converged Match Windows
0.15 T T T T T T T T T =] 0.35 zT T T T T T T T —
) e-—— 88-TD
oi X i 03 o 6x6- PD J
\ a——^ 16x16- 4)
\ ; a Sas A 14x14 -
V DX | 3——a 32x82 - TD d
\ A Lr A \ $ o 30x30 - PD à
€ A A Ae = > A $ 4
© k E x eee & = + 02h er ene
X 3 Beck eeu E eS x
& E REB. : =
2 E 2
t e S 05H ES à
2 2
a a
= £
$ 5 01H |
ui S |
c - e
— — c
à 7 7 3 0.05 | 4
6e——e 8x8- TD S LA SEE j
o. 0 6x6-PD N A ; P
-o15r ] lY a A a >
4———^ 16x16 - TD oH 14 ee t y MT ma AA TÉ 4
a ^ 14x14 - PD FX M gy
Fae -- E 4 /
-0.2- 32x32 - TD -0.05#; ¥
a o 30x30- PD
-0.25 1 L 1 1 L 1 L 1 1 -0.1 1 L L 1 1 L 1 L ep L EN
70 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80 90 100
Percentage of DCT Coefficients Taken in Least Squares Adjustment Percentage of DCT Coefficients Taken in Least Squares Adjustment
Figure 4: The effect of taking only a fraction of the available DCT coefficients in each least squares adjustment on the
mean error in the measure of a known disparity. The mean errors for the pixel domain algorithm are shown as dotted
lines for comparison. Results are shown for the “Redland” imagery (left) and the “Willunga” imagery (right).
6.3 Standard Deviation of the Disparity Errors
The standard deviation of the disparity errors represents the accuracy of the matching result, and in the case of the errors
having zero mean, which is approximately true for our data, it is equivalent to the RMS error in the disparity estimate,
Figure 5 shows that for all window sizes, and both images, the standard deviation of the errors starts off large, and as
more DCT coefficients are added, quickly reduces, and then flattens out, after which adding further DCT coefficients has
little impact on the accuracy. The knee in the curve occurred at around 5% to 10% for the 32 x 32 window, at around
20% for the 16 x 16 window, and around 25% to 30% for the 8 x 8 window. In all cases, after the knee, the accuracy was
comparable or better than that achieved by the pixel domain algorithm, sometimes markedly so.
Standard Deviation of Disparity Error for Converged Match Windows Standard Deviation of Disparity Error for Converged Match Windows
es T T T XT ER rate 14 T DENEN
= —À—,——Ó | f -
= | [
|e—o 88-TD | | | | ae, 83-10
| | | |
07 | 66-PD | el 66-PD | |
Q |4——^ 16x16 - TD | | FI La ^ 16x16-TD |
S \ ja a 14x14-PD | | 2 | AA 14x14 -PD |
$ 06- \ |9—9 32x32- TD | $ | |e——a 32x32- TD |
& |a a 30x330-PD|. à & 1n | © o 30x30 - PD |
$ LM 5 | L =
| = |
S | 8 0.8H
$9 | 2
041 = F | 5 ?
S | \ S esl
& ^1 ? 9|
$ 0.3 X | & 4|
a V aA |
© \ | $9 | |
$ N | 8 0.4} 1
S 0.2 x S : |
& Vr í ^ # ge m o
\ x 4 Ig |
A 0.2 £ |
s | | à.
93 FER 78— eo i eR EE a Eu d | RR S A A A 4
+ N x A es a £x
+ +
P 8
. —————— - LL "i -— 1 J o. A 1 A L L I 1 1 J —
0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80 90 100
Percentage of DCT Coefficients Taken in Least Squares Adjustment Percentage of DCT Coefficients Taken in Least Squares Adjustment
Figure 5: The effect of taking only a fraction of the available DCT coefficients in each least squares adjustment on the
standard deviation of the error in the measure of a known disparity. The pixel domain values are shown as dotted lines
for comparison. Results are shown for the “Redland” imagery (left) and the “Willunga ™ imagery (right).
6.4 Average Convergence Time
The average time for match windows to converge is shown in Figure 6. For the 8 x 8 window, the times for the DCT
domain algorithm are comparable to the pixel domain for DCT coefficient percentages up to about 30%, but then gradually
increase as further DCT coefficients are added. For the 16 x 16 window, taking between 10% and 30% of the coefficients
resulted in reducing the average convergence time to about 50% of the pixel domain time in one image, and about 75%
in the other image. The improvements were more pronounced for the 32 x 32 window, where in both images the average
convergence time was under 50% of the pixel domain time, when between 5% and 20% of the DCT coefficients wert
taken.
766 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000.
Figi
avel
hori
6.5
For
that
nun
spec
win.
alga
wer
resu
expx
bea
It w
mat:
32.
deve
inm
Mat
dom
also
filter
Su)
such
two
matc
This
fewe
over
redu
the I
the s
the s
coef
grea
All«
the s