The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B5. Beijing 2008 
820 
During structured light 3D scanning patterns of parallel light 
stripes are projected to the surface and a camera (or several 
cameras) captures the scene. From different viewpoints, the 
pattern appears geometrically distorted due to the surface shape 
of the object, and a pattern analysis and triangulation can 
recover the objects’ 3D coordinates. 
The accuracy evaluation of this scanner differed from those of 
previous methods, because the skull and the cadaver head were 
no longer available for us and a photogrammetric test-field was 
used instead (Figure 10.). The 3D model produced by the 
scanner combined from the overlapping point clouds of six 
scanning sequence, each of them covered a part of the test-field. 
The model coordinates has to be corrected by a scale factor. 
The resulted accuracy was characterized by the RMS error, 
which was 0.54 mm for the merged 3D model. Accuracy was 
about 20-25% better for each scanning part individually. This 
accuracy is considerable for our aims, but the multiple scanning 
procedure is not, because the complete immobility of the living 
subjects cannot be assured. Recently we also have been testing 
a scanner built for face scanning (Figure 11), which can 
generate 3D model in one step (approx. 4 seconds), but its 
accuracy is not yet calculated. 
Figure 10. Test measure with the 3D scanner 
measurements, because the common image matching 
algorithms seems to fail on the human skin as a texture. 
There is another way to reduce the amount of required 
workforce. Obviously, the less manual measurement needs less 
time and workforce, so we provided an estimate of the needed 
point density and the useful distribution of the measurement 
points. The scanning methods mentioned above produces point 
clouds of 2-300000 points per subject. Most of these points are 
unnecessary, because these points can be interpolated from their 
neighborhood accurately enough. Basically high curvature parts 
of the face require high point density, almost plain parts require 
low density. A filtering method has been defined to achieve 
these requirements. (Varga, 2008) The method based on a 
simple curvature estimation along the contour sections of the 
face model. Three consecutive points define an arc. when this 
arc is “straight enough”, the middle point was found to be 
unnecessary. The threshold was chosen in such a way that the 
RMS error of the distance between the original and the 
interpolated points had to be less than 1 mm. The number of 
unnecessary points was 68-95% of the original content, 
depending on the anatomical parts of the face. Figure 12 shows 
the original and the reduced point cloud. 
Figure 12. white/grey: measured points; black: undisposable 
points 
Figure 11. Face scanner 
2.4 Photogrammetry 
According to the literature (i.e. Fraser, 1996) and our previous 
experiences (i.e. Toth, 2005), we are nevertheless certain that 
photogrammetry can produce adequate 3D data by using a 
multi-station convergent photogrammetric network for face 
reconstruction. The most serious disadvantage of the 
photogrammetric method is the extreme time- and effort 
requirements of processing. Further investigation required to 
determine which part of the process can be automated (image 
matching, feature extraction etc.). We have been investigating a 
new image matching method based on colorimetric 
3. CONCLUSION 
The preliminary results of our work can be summarized as the 
followings: 
• A special material was developed for X-ray 
photogrammetric purposes. This resin is transparent 
both for the visible light and the X-ray, and suitable to 
have adequate marking in it. Using this test-field, 
numerical accuracy values were given to the usability 
of X-ray images and Direct Linear Transformation. 
• Accuracy of CT-generated 3D model was analyzed, 
and it suggests that this method is suitable for face