In: Paparoditis N., Pierrot-Deseilligny M.. Mallet C.. Tournaire O. (Eds), IAPRS. Vol. XXXVIII. Part ЗА - Saint-Mandé, France. Septeniber 1-3, 2010 
comparison of the results with the reference. The reference 
contains 85 segments. However, 15 of these segments represent 
chimneys and are smaller than 1 nr. None of these chimneys 
could be found, but of the remaining 70 planes, 58 (82.9%) are 
detected correctly (based on a visual inspection). Another 10 
planes (14.3%) are detected, but are split into two or more 
segments by our algorithm, and only three planes are missed. 
One of them is a small plane (1.5 nr) that belongs to a dormer, 
whereas the other two correspond to a small building structure 
that was missed in the detection and thus is outside the region of 
interest for segmentation. Eight planes are affected by 
topological errors in the sense that a part of one plane is 
erroneously added to another plane so that the neighbourhood 
relations are affected (the splitting of planes can also be 
considered a topological error). Most of them affect dormers 
that intersect the main roof plane: the part of the roof plane that 
separates the dormer from the ridge is assigned to the dormer. 
Most of these errors occur with the building used for 
Figures 1 - 5, which is the only building where planes had to be 
added to the segmentation based on the point cloud (cf. Section 
2.4). There are five false positive planes in the interior of the 
buildings. They are small structures at the transitions between 
several roof planes. There are four false positives outside the 
buildings. One of them is a parasol, the others are artefacts that 
correspond to vegetation and could possibly be removed based 
on their radiometric content. The delineation of the detected 
roof planes in Fig. 6 is not very precise yet. The deviations are 
in the order of magnitude of 3-5 image pixels and are clearly 
influenced by the resolution of the point cloud. In the future, the 
delineation of the roof planes shall be improved by integrating 
image edges in a way similar to (Rottensteiner et al., 2004). 
<7\Y 
0z 
\7 
LA£ - j min 
k 
N s 
a 
NP 
1УЛ mm 
N m 
Nol 
Qmin 
0.15 m 
0.075 m 
2.0 m 
2.0 
5 
1% 
4 
16 
5% 
33.3% 
Table 1. Parameters used in the experiments. The symbols are 
explained in the text. NP min is given for the first 
segmentation process. N OL is 5% of the number of 
points in either of the segments. 
4. CONCLUSIONS 
We have presented a new method for roof plane detection based 
on the segmentation of multiple aerial images and a point cloud. 
The method makes use of the observation that segmentation 
results differ over various images, so that a combination of 
several segmentations may result in a better separation of roof 
planes than a segmentation of a single image. The method was 
applied to detect the roofs of several complex buildings in a 
densely built-up historic town centre, based on images having a 
resolution of 8 cm and an ALS point cloud with an average 
point spacing of about 0.5 m. The results show that the method 
is capable of detecting most of the roof planes correctly with 
only a few false positives. The accuracy of the delineation 
corresponds to the average resolution of the point cloud and 
requires improvement for building reconstruction, e.g. by 
integrating image edges. The LoD of the segmentation is 
restricted by the resolution of the point cloud, because a 
minimum of three points is required for an image segment to be 
considered a candidate for a plane. This could be improved by 
integrating 3D edges derived from multiple-view matching, 
because only one further point (or an additional 3D edge) would 
be sufficient to support a plane hypothesis. It has to be noted 
that the current procedure essentially works in 2.5D because the 
projection of the segmentation results into the (Ah’) plane in 
object space is used for segment matching and because the 
combined segmentation is also represented in the (Ail) plane. 
This restriction could be overcome by defining and matching 
the planar segments directly in image space, in which case the 
images would be connected via the object planes. However, this 
would require a specific visibility analysis for any pair of 
images involved in the process. Finally, the method still needs 
to be verified for point clouds generated by image matching. 
ACKNOWLEDGEMENT 
The Vaihingen data set was provided by the German 
Association of Photogrammetry and Remote Sensing (DGPF; 
Cramer & Haala, 2009). More information can be found at 
http://www.ifp.uni-stuttgart.de/dgpf/DKEP-Allg.html (German) 
REFERENCES 
Ameri, B., Fritsch, D., 2000. Automatic 3D building 
reconstruction using plane-roof structures. In: Proc. ASPRS 
Spring meeting, on CD-ROM. 
Briigelmann, R„ Forstner, W„ 1992. Noise estimation for color 
edge extraction. In: W. Forstner, S. Ruwiedel (eds.), Robust 
Computer Vision, Wichmann. Karlsruhe, pp. 90-107. 
Cramer, M., Haala, N., 2009. DGPF Project: Evaluation of 
digital photogrammetric aerial based imaging systems - 
overview and results from the pilot centre. In: JAPRSIS 
XXXVIII (1-4-7/W5), on CD-ROM. 
Dominger. P., Nothegger, C., 2007. 3D segmentation of 
unstructured point clouds for building modelling. In: IAPRSIS 
XXXVI (3/W49A), pp. 191-196. 
Drauschke, M., 2009. An irregular pyramid for multi-scale 
analysis of objects and their parts. In; 7 th 1APR-TC-15 WS on 
Graph-based Representations in Pattern Recogn., pp. 293-303. 
Drauschke, M., Roscher, R., Labe, T., Forstner, W., 2009. 
Improving image segmentation using multiple view analysis. In: 
IAPRSIS XXXVIII (3/W4), pp. 211 -216. 
Forstner, W„ Wrobel, B., 2004. Robust estimation. In: Me 
Glone (ed.), Manual of Photogrammetry. 5 th ed., American Soc. 
Photogrammetry & R. S., Bethseda (Md), USA, pp. 103-110. 
Haala, N.. Hastedt. H., Wolff, K., Ressl, C., Baltmsch. S., 2010. 
Digital photogrammetric camera evaluation - generation of 
digital elevation models. PFG 2(2010) (in print). 
Khoshelham, K., 2005. Region refinement and parametric 
reconstruction of building roofs by integration of image and 
height data. In: IAPRSIS XXXVI (3/W24f pp. 3-8. 
Rottensteiner, F„ Trinder, J., Clode, S., Kubik, K., 2004. Fusing 
airborne laser scanner data and aerial imagery for the automatic 
extraction of buildings in densely built-up areas. In: IAPRSIS 
XXXVI (B3), pp. 512-517. 
Rottensteiner. F., Trinder, J., Clode, S., Kubik, K., 2005. 
Automated delineation of roof planes in LIDAR data. In: 
IAPRSIS XXXV (3/W19), pp. 221-226. 
Song, Y., Shan, J., 2008. Building extraction from high 
resolution color imagery based on edge flow driven active 
contour and JSEG. In: IAPRSIS XXXVII (B3A), pp. 185-190. 
Vosselman. G., Dijkman, S., 2001. 3D building model 
reconstruction from point clouds and ground plans. In: IAPRS 
XXXIV (3/W4). pp. 37—42. 
250