tanbul 2004 
— —— 
  
es of points; 
tions of the 
ng between 
| portion of 
m), by the 
/o 1n almost 
t where the 
ollect data 
points). In 
f data, also 
grammetric 
ere used as 
:gistrations 
processing 
to manage 
ation time. 
me. At the 
ing to the 
que model. 
  
  
  
  
national Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B5. Istanbul 2004 
Inte ) ; : d / l 
  
  
Figure 4. Surface model with 
  
  
  
  
  
: & different levels of noise and 
max distance avg distance|st deviation smooth filter: it is possible to 
e mi ini m. appreciate different details in 
0.003 0.001 00006 | the material texture. 
j Table 2. The shifting 
—-| 0.004 0.001 0.001 : ; 
applied to the points after 
— | different levels of noise and 
0.008 2,002 9.004 smooth filter 
  
  
  
  
After the global registration, the final residuals are, on average, 
lower than one centimeter; this can be considered a good result 
related to the accuracy of the employed instruments (= 6 mm). 
3.2.2 Preprocessing: Often during the scanning it is not possible 
to remove the obstacles on the scene: in Perugia we could not 
avoid scanning elements as scaffoldings and other materials of 
the yard in progress in addition to the weeds. Sometimes these 
obstacles give great shadow which have to be considered in the 
survey project. Automatic selection with a distance based filter, 
is only useful in preliminary, approximate cleaning of data. In a 
lot of parts, manual selection was used to obtain more accurate 
results. The reduction of the points was considerable (14%). 
The final cloud of points (17 million points) was partitioned 
marking the boundary of relevant architectonic portions of about 
one million points in order to optimize the following operation 
and to allow data management in real time. These clusters of 
data were elaborated and joined one by one in a unique surface 
model, drastically decimated. Saving the data in every step of 
the elaboration would allow to assemble the model in every phase 
of processing. 
Noise reduction: The noise reducing operation was carried out 
by a filter available in Raindrop Geomagic: using statistical 
methods, the operation determines where the points should lie, 
then moves them to these locations. Depending on the magnitude 
of the errors, it is possible to choose a minimum, medium, or 
maximum noise reduction setting. 
Two options help optimize the operation for the type of model 
with which it is working. If the point set represents a freeform or 
organic shape, the operation reduces the noise with respect to 
surface curvature. If it is a mechanical or prismatic shape, the 
operation helps keep features sharp such as edges. After the noise 
reduction is complete, statistics are displayed in the Dialog 
Manager that indicate the Maximum Distance, Average Distance, 
and Standard Deviation of the points from their original positions. 
Tests with range maps at a different resolution were performed: 
the first one with a single range map, acquired at maximum 
resolution (6 mm) on the capitals area and the second one with a 
lot of range maps acquired at 1.5 cm of resolution on a more 
extended portion of the transept. 
  
First we tested the smoothness level parameter on the range map 
carried out with maximum resolution. In function of the obtained 
results, summed up in the table, we chose to apply to all the data 
a noise reduction with medium smoothness level. In more realistic 
operating conditions, however, there are a lot of range maps 
acquired with less resolution. In these cases, the effects of the 
overlapping add to the noise effects and join themselves. We 
have also noticed that the combined application of noise and 
smooth filters involve a significant reduction of the descriptive 
capability of the model in order to represent both the surface 
texture and the edges of the architectonic elements. At the end, 
we preferred to apply a noise reduction with a minimun 
smoothness level to the final model. 
Decimation: Data derived from laser scanning are characterized 
on one hand by redundancy of measured points and on the other 
hand by no critical selection to describe the morphology of the 
object. The decimation procedure is aimed at reducing the huge 
number of points in order to give a better approximation to the 
shape of the object. It is possible to use different criteria: 
- random sampling, a percentage decimation, applied to the whole 
cloud of points in a random way; 
- uniform sampling, that subdivides the model space into equally 
sized cubical cells (the dimension of the cells is a function of the 
fixed level of decimation) and deletes all but one point from 
each cell; 
- curvature sampling, in which points that lie in a high curvature 
region remain in order to mantain the accuracy of the surface 
curves; because flat regions require less detail, points in those 
regions are more likely to be deleted. On the same range map 
above mentioned, we applied different decimation algorithms: 
to apply random sampling is similar as to acquire data with a 
wider sampling step, useful only for coarse decimation; the 
uniform sampling allows to have more regular triangles in the 
surface but the descriptive capability of the complex shape is 
  
Figure 5. 
Curvature sampling. We can 
note the effects of the 
decimation: stronger in the 
regular surfaces 
j