CI PA 2003 XIX th International Symposium, 30 September - 04 October, 2003, Antalya, Turkey 
BASTA and documented in Kalisperakis & Tzakos (2001), has 
been primarily intended as an educational tool. It functions to 
tally within a commercial CAD environment, exploiting its tools 
to allow convenient image measurement as well as editing. 
Figure 2. The five images of the ‘Tsopotos’ residence. 
The program solves the bundle without control (‘relative solu 
tion’); it offers options for both full and partial self-calibration 
(camera constant, principal point and/or radial lens distortion); 
it may handle control coordinates as observations; further, it can 
accept any of the three geodetic coordinates of control points as 
known and the others as unknowns, a very useful feature in the 
context of architectural photogrammetry (see 4.2). For both pro 
grams the same 92 tie points were used, but image coordinates 
were measured independently within the two environments. An 
average of 3.7 intersecting rays per object point formed a strong 
bundle configuration. The image at the bottom of Fig. 2 has a 
larger scale. Adjustments for all 5 images but also for the re 
maining 4 have been carried out to check for any possible diffe 
rences due to this particular image (Rova, 2003). 3 
3. ADJUSTMENTS WITHOUT CONTROL 
As no particularly reliable control coordinates were at hand (see 
4.1 below), the first step was to solve the bundle without geode 
tic control, i.e. in arbitrary systems, properly scaled in the case 
of BASTA whose results were considered as reference data re 
garding PhotoModeler. Outcome of the adjustments, which re 
lied on the assumed ‘nominal’ values for inner orientation para 
meters c = 80 mm and x<, = y 0 = 0, is 92 tie point coordinates in 
the two systems of BASTA and of PhotoModeler (B and PM, re 
spectively). Point sets represent shape reconstruction and can be 
compared to each other by means of a 3D similarity transforma 
tion (Rova, 2003). The overall standard error (ct s ) of this trans 
formation in all three axes describes the ‘closeness’ of results 
from the two programs. 
Other measures of precision are also available. For BASTA, this 
is the standard error of image coordinates (g 0 ) and, further, the 
overall standard deviation (ct t ) of tie points, resulting from the 
variance-covariance matrix of unknowns in the bundle adjust 
ment. On the other hand, PhotoModeler produces a ‘tightness’ 
value (t) for every tie point, representing the maximum distance 
among intersecting rays as percentage of the largest object di 
mension. Furthermore, PhotoModeler can perform what it calls 
a ‘self-calibration’ (from now on PMS), meaning a small ‘inter 
nal’ adjustment of camera parameters to ‘optimize’ the solution; 
however, users do not actually see it. Results from the two pro 
grams are concentrated in Table 1. In Table 2 results from the 
3D similarity transformations are given. 
Table 1 
Precision estimations from 92 tie points 
B 
PM 
PMS 
CT 0 (Jim) 
18.3 
4 images 
0> T (cm) 
1.3 
t (cm) 
3.8 
4.0 
CT 0 (ftm) 
19.2 
5 images 
G T (cm) 
1.5 
t (cm) 
3.3 
3.2 
Table 2. Accuracy g s (cm) of the 
3D similarity transformations 
B-PM 
B- PMS 
4 images 
2.0 
1.9 
5 images 
2.0 
1.9 
Image residuals (represented as G 0 = 1.3 pixel) and tie point pre 
cision G T are acceptable if the poor quality of object points, due 
to building decay, is taken into account. ‘Tightness’ values from 
PhotoModeler (being larger than G T by a factor 2 to 3) appear to 
be grossly comparable to Ctt since they refer to the largest devia 
tions among rays. Finally, the similarity measures g s in Table 2 
between point sets confirm that reconstructed shapes are practi 
cally equivalent, within the precision of the two methods. 
4. SELF-CALIBRATING ADJUSTMENTS 
Besides defining the object system, the use of control points can 
also allow a self-calibrating approach to optimize reconstruction 
(and, additionally, provide information about camera geometry). 
In the present case, separate old elevation drawings prepared by 
students at the School of Architecture of NTUA were available. 
A few control points, unfavourably distributed but sufficient for 
self-calibration, were extracted graphically. Being of unknown 
accuracy, however, the plots were regarded as unreliable (a fact 
also established in the experiments). Under these circumstances, 
the next tests were essentially intended to check the camera cali 
bration processes rather than the accuracy of space coordinates. 
4.1 Use of control points 
PhotoModeler does not allow recovery of lens distortion (but it 
accepts existing data about it). After some experimenting, it was 
concluded that, for 5 images, camera constant (c) and principal 
point (x 0 , y 0 ) could be determined from 7 control points (at least 
5/image). For 4 images, 6 points were required on all images (5 
points/image did not suffice for the principal point); tests were 
made with 7 control points, too. The same control was also used 
for self-calibration in BASTA (both with and without estimation 
of radial lens distortion). Being theoretically in the same object