Mean values have been computed on three outputs :
e residuals on points (see Figure 7)
e residuals on segments (see Figure 8)
e main axis of orientation error ellipsoid (see Figure 9)
We do not present location error ellipsoid parameters, since bad
spatial configurations, especially occurring with few segments
and points, returned wrong values. With few observations, our
approximation which enables error propagation is not valid any-
more.
Mean residuals on points goes down to less than 0.4 pixel. Resid-
uals on segments are also satisfactory. There is still a weighting
between primitives issue, which is not yet overcome.
Image residuals, pixel
eo
N
1
94] | 3 points
J | ^ 12points
0,0 4
T T T T T T T T T T T T T
0 ad: 2 23.1454 5:,.61.7. 8 9110 110
Number of segments
Figure 7: Image residuals on points.
0,6 4
3 o5]
a
Sd
S 04. Hd
E x
> 7 X
o
c 034 X X
o
o | xx
S
9.02
© X
= uate asset
2 0,1 x Segments only |
E Segments + 3 points |
7 > Segments + 12 points |
0.0 ds N EE
1 T T T T T T T T T T T T
0 1 2 3 4 5 6 7 8 9549 11 42
Number of segments
Figure 8: Image residuals on segments.
First results show that using few segments (less than 6) is quite
dangerous (it is a little bit obvious). Much interesting is the fact
that combining 3 points with segments seems to be quite equiva-
lent to taking 12 points with segments (even if the scale represen-
tation in y axis is good).
Some results are presented through figure (10) and (11). Here,
each 3D point is drawn with its corresponding image radiometry.
These 3D views highlight pose accuracy, especially on radiomet-
ric discontinuities.
1134
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004
X
x Only segments
94 Segments + 3 points
> Segments + 12 points
"S
=
> 3-
o
e
x
©
ec
= 2
E
S x
o
uo
& 14
©
5 -
ul : A nien Nocmods
0 T T T T T T T T. T T T T
e
—_
N
C
>
C
o
7-78 49: 40 14-12
Number of segments
Figure 9: Relative effects of features on orientation estimate.
Figure 10: Results on the Vincennes castle chapel. Here, we show
the point cloud, where each point has been assigned its image
corresponding RGB pixel.
Intei
and
age
in ri
thot
a hi
met
few
Seg
feat
the
SES.
feat
Err
cur:
on ]
Thi
tion
has
mel
Fur
loo
ing
sul
ine
Fur
pro
ing
fing
To
par
tun
wit
WO
Fir
titi
acc
me