ISPRS Commission III, Vol.34, Part 3A ,Photogrammetric Computer Vision“, Graz, 2002
block model obtained from converting 2 1/2 dimensional
GIS (geographic information system) data. An overview
of a part of our 3D block model can be seen in Figure 1.
The 3D block model is augmented using image sequences
captured by a digital consumer camera from arbitrary po-
sitions. Currently, we use a Canon D30 with a geometric
resolution of 2160 x 1440 and 12bit per pixel radiomet-
ric resolution. The images have to be captured using short
baselines, thus a digital video camera with a reasonably
high resolution will work as well. Our work flow consists
of five consecutive steps which will be explained in the fol-
lowing subsections.
3.1 Line Extraction and Vanishing Point Detection
The line extraction stage starts with an edge detection and
edge linking process and yields contour chains with sub
pixel accuracy. For all contour chains of sufficient size a
RANSAC [3] based line detection method is applied. Pairs
of contour points are randomly picked from the contour
and a potential line segment is formed. For this line seg-
ment inlier points, that are points with a small perpendicu-
lar distance to the segment are searched. The line segment
with most inlier points is considered the best hypothesis for
the line. The final line parameters are estimated by a least
squares adjustment over the inlier points. A final grouping
process merges collinear segments that lie close to each
other.
The vanishing point detection is based on the method pro-
posed by Rother [9]. In this approach the previously ex-
tracted line segments are used for the detection. Each in-
tersection is treated as potential vanishing point and the
weight for the intersection is determined by testing against
all other line segments. The smaller the angle difference
between a line segment and the vector pointing from the
mid point of the segment to the intersection, the higher
the contribution to the weight of the accumulator cell for
the potential vanishing point. If the angle difference is too
large, the weight of the accumulator cell is not increased.
The intersection with the maximal weight is then reported
as vanishing point.
3.2 Relative Orientation of Image Pairs
We developed two completely different methods to cal-
culate the relative orientation of image pairs. The first
method is based on calculating corresponding points within
an image pair. Therefore an area based hierarchical matcher
is used. In the second approach vanishing points are used
to solve the relative orientation problem without the neces-
sity to perform time consuming point to point correlation
in image pairs.
3.2. Relative Orientation from corresponding points
In order to determine the relative orientation of an image
sequence, we need to find corresponding points in all adja-
cent image pairs. In our approach we focus on an iterative
and hierarchical method based on homographies to find
this corresponding points inspired by a work published by
Redert et al. [8]. For each input picture an image pyramid
A - 188
is created and the calculation starts at the coarsest level.
Corresponding points are searched and upsampled to the
next finer level where the calculation proceeds. This pro-
cedure continues until the full resolution level is reached.
This hierarchical method convergences fast and avoids lo-
cal minima solutions especially when having repetitive struc-
tures within a facade.
(a) First input image
(b) Second input image
(c) Visualization of our accuracy measure; Crosses indi-
cate corresponding points used for the calculation of the
relative orientation.
Figure 2: Input images and visualization of our accuracy
measure.
A reliable calculation of the relative orientation can be done
using a set of corresponding points which should fulfill
some properties. They have to be well distributed over the
images with a good location accuracy and a low outlier
rate. To achieve this requirement it is necessary to calcu-
late an accuracy measure for all calculated corresponding
points. This is done during the matching process and is
calculated within a cost function based on the distribution