CMRT09: Object Extraction for 3D City Models, Road Databases and Traffic Monitoring - Concepts, Algorithms, and Evaluation
edges should very well coincide with a strong image line.
However, in practice there are often a small offset and an
angle between a model edge and its corresponding image
line. The optimal angle and distance threshold value are de
pendent on the quality of exterior orientation parameters and
focal length.
2. A best match is chosen from all candidates according to
either the collinearity of the candidates or the candidate’s
length (see the purple lines in Figure 5). It is a common
case that a strong line is split to multiple parts by occlusions
or shadows. If a number of hough line segments belong to
a same line, we set this line as the best match. If not, the
longest candidate is chosen as the best match.
Figure 5: Matching model edges with image lines (Blue: model
edges’ projection in the image; red: Hough lines; green: candi
dates; purple: the best matches)
No spatial index is established in the image space to improve the
comparison efficiency, because the search space is already local
ized to a single building facade, which includes only dozens of
edges and Hough lines.
A limitation of this matching method is that it can hardly de
termine the correct corresponding edge if too many similar line
features are within searching range. Simply comparing the geom
etry properties of position, direction and length are not sufficient
in this case. For example, the eaves in Figure 5 result in many
significant lines and they are all parallel and close to the wall’s
upper boundary edges. These eave lines can be distinguished if
the eave is also reconstructed and included in the facade model,
but ambiguity caused by pure color pattern is still difficult to be
solved.
5.3 The refinement strategy
After matching, most model edges should be associated with a
best matched image line. These model edges are updated by pro
jecting to their best matched image line. There are some model
edges which don’t match any image lines. If no change is made
to an edge with its previous or next edge changed, strange shapes
like sharp corners and self-intersections may be generate. There
fore interpolations of the angle and distance change from the pre
vious and next edges, are applied to the edges without matched
image lines. With these refinement strategies, an original model
is updated to be consistent with the geometry extracted from im
ages, and the model’s geometry validity and general shape are
also maintained.
Finally, the refined model edges in image space need to be trans
ferred back to the model space. Because the model edges are only
moved on their original 3D planes, which is known, the collinear
ity equations are used again to calculate the new 3D positions of
all the modified model vertices.
6 TEST CASES
In this section, three data sets are experimented with the presented
refinement method. The building models are produced with the
reconstruction approach introduced in Section 3. All the images
are originally provided as Cycloramas. The central perspective
conversion and exterior orientation calculation follow the pro
cesses explained in Section 4.
6.1 The restaurant house
The inconsistencies between the model edges and image lines in
Figure 6 are mainly due to inaccurate exterior orientation of the
image. It is difficult to pick an image point accurately by man
ual operation. Picking the corresponding point in a laser point
cloud is also a difficult job. Automated texturing of building fa
cade models is desired in the context of our research. The quality
of the exterior orientation is a key issue to the texturing effect.
Even a minor inaccuracy in the exterior orientation parameters
can lead to poor texture result, as shown in Figure 7(a). Apply
ing our refinement method, several model edges are linked with
their matched image lines (see Figure 6(b)), and are updated ac
cordingly. The texture result is significantly improved as shown
in Figure 7(b), with the sky’s background color removed. How
ever, the middle top part of the facade model is still not refined,
because this image part is too blurred to output a Hough line.
Figure 6: Matching model edges with image lines for refining the
restaurant house’s model
6.2 The town hall
The upper boundary of the town hall in Figure 8(a) contains a
lot of tiny details, which are well recorded by laser scanning and
modeled as sawtooth edges in the building facade model. Instead
of adjusting the outline generation parameters in the reconstruc
tion stage, we can also use the presented image based refinement
to smooth the model outline. Figure 8(b) shows the matching
step. The model’s upper edges are successfully matched to the
strong lines, which actually come from the eave. In this example