Boehm, Jan 
  
  
  
  
  
  
(a) adapter (b) bigwye 
Figure 6: Shaded view of the CAD models. 
  
Figure 7: Results on test dataset “adapter” at different poses. 
computed. If the CAD feature F contains more than one HK sample, such as {(H,, K1),..., (H,, K. n)}, the distance of 
the feature vector p to feature F in HK space is determined as d(p, F) = min(d(p, (H;, K i))). The pixel is assigned the 
surface label of the closest feature. The assignment is thresholded to prevent severe misclassification. Pixels which were 
not assigned any label remain unclassified. As it has been noted above, this is not an object recognition process. Only 
the features for object recognition are being extracted. For example several planar patches may be present in the scene. 
They will all be assigned the same label. After the classification connected component analysis is performed. This step 
assigns different labels to unconnected patches of the same surface type. The resulting regions define the scene features 
fi. Searching techniques are then needed to resolve the ambiguities for the object recognition process. 
4 EXPERIMENTS 
First experiments were made based on data from the SAMPL database maintained by P. Flynn (Campbell and Flynn, 
1998). The data base provided us with CAD models and range images for various simple parts. Most of them contain 
only cylindrical and planar surfaces and have few surfaces (less than 10). We chose two objects from the database, the 
first dataset is called "adapter". It is a simple combination of two cylinders 6(a). 
For the computation of curvatures from the range image we had to chose the size of the neighborhood for the surface fit. 
Clearly, choosing the mask size is a trade-off between reliability and accuracy near edges. When choosing a small mask 
curvature computation will be strongly affected by noise, due to the small number of points considered for regression. 
c 
On the first object we used a mask size of 7x7. Figure 7 shows the res 
the contours of the extracted regions. As we can see the algorithm performs well and gives sharp 
edges. But as this is a very simple object the curvature characteristics of the three surfaces containe 
distinguished easily and curvature estimates did not have to be extremely precise. 
The second object, more complex, contains 5 cylinders and the according planar surfaces. Figure 6(b) shows this d 
called “bigwye”. Here more surfaces which are quite similar have to be distinguished. We needed a better estim: 
curvatures and therefore chose a mask size of 15x15. Figure 8 shows the results. We can see the effect a | 
has on the performance of the algorithm near edges. Surface boundaries are not as sharp as before. But 
results are quite encouraging as the surfaces can be reliably detected independent of the pose of the object 
arge mask size 
still the overall 
For the second experiment we used a part of more complex shap 
e. For our research work we created a test object 
ourselves. Using Pro/ENGINEER we created a model containing se 
veral different surface types expected in an industrial 
  
  
80 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B5. Amsterdam 2000. 
  
object ii 
from an 
obtaine« 
stripes. 
size was 
precise 
small su 
great re 
to that c 
features 
5 CO 
We hav 
estimate 
minimu 
used to 
corresp 
CAD sy 
which a 
model t 
6 ACI 
The wo