anbul 2004 
WaveScan 
ımetry and 
hanics and 
EB), both 
he Federal 
1e German 
funding 
the advice 
ny Water 
€) Norden 
ney of the 
  
AUTOMATIC CORRESPONDENCES FOR PHOTOGRAMMETRIC MODEL BUILDING 
Dr. Gerhard Roth 
Computational Video Group, Institute for Information Technology 
National Research Council of Canada 
Gerhard.Roth@nrc.ca 
Working Group TS ICWG V/III 
KEY WORDS: Photogrammetry, Modelling, Matching, Automation, Feature, Semi-Automation. 
ABSTRACT 
The problem of building geometric models has been a central application in photogrammetry. Our goal is to partially 
automate this process by finding the features necessary for computing the exterior orientation. This is done by robustly 
“ computing the fundamental matrix, and trilinear tensor for all images pairs and some image triples. The correspondences 
computed from this process are chained together and sent to a commercial bundle adjustment program to find the exterior 
camera parameters. To find these correspondences it is not necessary to have camera calibration, nor to compute a full 
projective reconstruction. Thus our approach can be used with any photogrammetric model building package. We also 
use the computed projective quantities to autocalibrate the focal length of the camera. Once the exterior orientation is 
found, the user still needs to manually create the model, but this is now a simpler process. 
1 INTRODUCTION 
The problem of building geometric models has been a cen- 
tral application in photogrammetry. Our goal is to make 
this process simpler, and more efficient. The idea is to au- 
tomatically find the features necessary for computing the 
exterior orientation for a given set of images. The claim is 
that this simplifies the model building process. To achieve 
this goal it is necessary to automatically create a reliable 
set of correspondences between a set of images without 
user intervention. 
It has been shown that under certain conditions it is pos- 
sible to reliably find correspondences, and to compute the 
relative orientation between image pairs. But these con- 
ditions are that the images are ordered, and have a small 
baseline, such as those obtained from a video sequence. 
However, in photogrammetric modeling the input images 
are not ordered and the baseline is often large. In this case 
it is necessary to identify which image pairs overlap, and 
then to compute correspondences between them. 
Our solution to this problem contains a number of inno- 
vations. First, we use a feature finder that is effective in 
matching images with a ^vide baseline. Second, we use 
projective methods to verify that these hypothesized matches 
are in fact reliable correspondences. Third, we use the pro- 
jective approach only to create a set of correspondences, 
and not to compute a projective reconstruction. At the end 
of the process these correspondences are sent to a bundle 
adjustment to find the exterior camera parameters. This 
makes the generation of the correspondences independent 
of the bundle adjustment. Therefore our approach can be 
used with any photogrammetric model building package. 
The computed projective quantities are the fundamental 
matrix between image pairs, and the trilinear tensor be- 
tween image triplets. The supporting matches for the ten- 
sors are chained across images that are found to be adjacent 
to create correspondence chains. We show that this process 
creates very reliable correspondence chains. Finally, we 
  
use the computed projective information to autocalibrate 
the camera focal length. 
Working in projective space has the advantage that camera 
calibration is not necessary. The other advantage is that it 
is possible to autocalibrate certain camera parameters, in 
particular the focal length, from the fundamental matrices. 
The complexity of computing the fundamental matrix, and 
trilinear tensor is less than that of the bundle adjustment. 
2 DISCUSSION AND RELATED WORK 
Model building is a very common industrial Photogram- 
metric application. The basic methodology has been un- 
changed for many years; calibrate the camera, choose the 
2D projections of the 3D points of interest in different im- 
ages, and then run the bundle adjustment. Then take the 
computed 3D data points, and join them together to cre- 
ate the model topology. The process is very manual, but 
has the advantage that the user can control the exact geom- 
etry and topology of the final model. Recently there has 
been some important recent work in the area of automated 
model building. However, the problem of finding the ap- 
propriate geometry and topology is a difficult and unsolved 
one. Thus the photogrammetric approach to model build- 
ing will likely be used for a considerable length of time. 
However, while it is very flexible it is also very labor in- 
tensive. 
Is it possible to partly automate the photogrammetric model 
building process and still leave it's flexibility intact? We 
believe the answer is yes. Normally the feature points used 
to build the model are also the ones used to run the bundle 
adjustment. But what if some other automatic process had 
already computed the exterior orientation? Then comput- 
ing the required 3D data points for the model would require 
only triangulation, instead of the bundle adjustment. The 
triangulation process is inherently simpler than the bundle 
adjustment, so this would simplify the process of creating