International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004
Input Data ^ Imase Sequence - —— ——— Lr
A 7i.
3D "CAD-Model"
of Planar Surfaces
Corresponding
RR
Self calibration
Calibrated Cameras
of Pairs
!
[ ;
| Rectification
I
I
Epipolar Geometry
r B» Guided
Correlation
Dense Disparity Maps
Depth Maps
I
I
I
I
I
Image Pairs in
I
I
I
Elevation Grid ;
|
|
|
I Creation of
l
|
|
I
Texturing of :
the Relief
Textured Elevation Grid
Figure 2: Overview of the processing chain from a coarse
to a refined model.
very generic structural models (e. g. several smooth planes
at different depth levels) are required. On the other hand,
different objects or parts of them are not separated by the
model and can not be treated individually by the visualiza-
tion process.
The other approach is to use models of objects which are
expected to be present in the scene like e. g. buildings in
an urban environment. This typically results in scene ob-
jects represented by polyhedral models made from plane
surfaces. For visualization, areas of the original images are
extracted and mapped as flat textures on each of the sur-
faces. The advantage is that the objects are already hierar-
chically structured entities that are easy to handle. But re-
lief structures within each surface are lost if they are — and
this is the general case — not contained in the model.
Both methods produce models which are less realistic on
close inspection and very oblique views because no relief
structure is visible. In our approach, a hybrid method that
combines the advantages of both reconstruction techniques
in a hierarchical manner is proposed (Fig. 1). Polygonal
boundaries of planar surfaces are taken as input and con-
strain the reconstruction of the finer details which are re-
trieved via dense matching. The result is a dense elevation
grid that can be used to replace the flat texture. Because the
elevation grid is restricted geometrically to the polyhedral
model at its borders, it fits to other surfaces without any
error.
2 PROCESSING CHAIN
In this section the processing chain from input data to the
final product is described. Refer to Fig. 2 for an overview.
2.1 Input Data
The inputs of the processing chain are an initial coarse ge-
ometric model of the object and a set of associated images
mapping the building. The coarse model is a wire frame
model of the building, in which planar surfaces are de-
scribed by 3D outline polygons. The images are needed for
both relief reconstruction and texturing. They are linked to
the polygons through known image coordinates of the ver-
tices. The process needs the determination of the camera
parameters by using the 3D coordiates. Therefore, it must
be assured that the 3D points are not coplanar in space so
that camera parameters can be computed for the 3D re-
construction. This is not the case for e. g. a single facade,
but can be circumvented by inclusion of some additional
points outside the plane. Here are two possible methods to
obtain suitable input models:
2.1.1 Retrieve model from images For a site where
only images exist, i.e. where no model has been gener-
ated or made available, it is possible to retrieve the wire
frame model from the images. Corresponding points in dif-
ferent images generally provide enough information to cal-
ibrate the cameras and to reconstruct the 3D coordinates of
the imaged points. Next, either the 3D bounding lines of
the surfaces are extracted and intersected or the corners of
these surfaces are connected in order to create the poly-
gons.
2.1.2 Coregistrate existing model to images We have
assumed that there already exists a wire frame model.
Here, the 3D points are already given and only have to
be marked in the images so that the corresponding image
coordinates are available. No knowledge about the cam-
era parameters is needed as input, but for later self calibra:
tion it should be known which of the parameters (e. g. focal
length) can be considered constant.
2.2 Self calibration
The estimation of a depth map requires knowledge about
the pose of the cameras as well as their calibration param-
eters. If these are not known they have to be computed
from the given point assignments. Such a task — simulta-
neous computation of inner and outer camera parameters
when no initial values are known — is commonly referred to
as auto or self calibration (Hartley and Zisserman, 2000).
In the case that corresponding 2D and 3D points are avail-
able, the following two step strategy could be applied:
(a) Linear Resection Resection computes the homoge-
neous 3x4 projection matrix P from corresponding image
points x; and world points X; which are related by
x; = PX; (1)
Using the vector cross product
x; x PX; =0 (2)
Eq. 1 can be transformed into an equivalent equation
Ap =0 (3)
Inter
—
Figi
out]
whe
the
tor.
(Ha
An
ima
pare
are
are
resu
a lii
bee
torti
erro
rica
(b)
bun
equ:
tern
are
The
kno
sum
sure
The
ters
rela
the :
23
Rec
sucl
and
COO!
ing
the
slan