tance of 
"ned the 
ch were 
s. Only 
e scene. 
his step 
features 
Flynn, 
contain 
ase, the 
face fit. 
ll mask 
ession. 
k sizes 
' of the 
id with 
to the 
can be 
lataset 
for the 
sk size 
)verall 
object 
ustrial 
  
Boehm, Jan 
  
  
  
  
Figure 8: Results on test dataset “bigwye”. Surfaces can be detected independent on the pose of the object. 
  
(a) shaded view of CAD model (b) the part in the milling machine 
Figure 9: Self-created model / part for experiments. 
object including among others planar, spherical, cylindrical, and free-form surfaces. This model was then manufactured 
from an aluminum block using a high precision milling machine. The size of the part is approximately 20 cm x 15 cm. We 
obtained range images from the object using our ABW stripe projector capable of projecting at a resolution of 640 x 640 
stripes. Our camera system is a Basler A113 at a resolution of 1300 x 1030 pixels. Approximately only half of the image 
size was used. Based on the experiences from the first experiments we used again a mask size of 15x15. This allows for a 
precise estimate of curvatures but it prohibits us to detect small details of the part. As we can see in the results of figure 10 
small surface details on the left of the object could not be extracted. However, large surface patches can be detected with 
great reliability. Figure 11 shows that the algorithm is capable to achieve a level of abstraction from the data comparable 
to that of the CAD data. These results are very encouraging for the matching stage where the correspondence of extracted 
features to those of the CAD model is to be established. 
5 CONCLUSION 
We have demonstrated a system detecting surfaces in a range image independent of the pose of the object. Reliable 
estimates of surface curvature are obtained from range images using a least squares surface fitting algorithm. A simple 
minimum distance classification has been shown to be adequate for range image classification when a CAD model is 
used to derive the ’master’ classes for the classification process. The surfaces extracted from the range image closely 
correspond to those of the CAD model. The procedure is able to process all surface types which can be expressed in the 
CAD system including critical curved surfaces and free-form surfaces. This is an improvement over previous systems 
which are restricted to certain surface types. These results are an important step towards our goal of establishing a CAD 
model based object recognition system for industrial parts which is able to process arbitrary objects. 
6 ACKNOWLEDGMENTS 
The work presented in this paper was supported by the German Research Foundation under research grant DFG-SFB514. 
  
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B5. Amsterdam 2000. 81