Semi-automatic teeth segmentation in 3D models of dental casts using a hybrid methodology - Emerging Trends in Image Processing, Computer Vision, and Pattern Recognition - page 363

2.1 3D Model Acquisition
The Universidad Nacional de Colombia Sede Manizales has a 3D digitizer VIVID 9i Konica
( Figure 1 ) . This scanner produces range images which constitute a valuable source of inform-
ation. Since range images cover the object's geometry from a specific point of view, several
shots are needed in order to reconstruct a whole model without occlusions.
FIGURE 1 Konica Minolta VIVID 9i 3D digitizer.
The views acquired with the range scanner must be aligned up into a single coordinate
space. This procedure is called registration (e.g., in Figure 2 ) .
FIGURE 2 Registration.
Once the views have been registered, the integration process starts. The goal of integration
is to generate a well-defined mesh or data set using the information coming from all the views
(partial meshes or point sets) captured during scanning ( Figure 3 ) . Furthermore, this process
seeks to eliminate redundant information in regions with litle variation in the surface, and to
ill small holes in the surface.
FIGURE 3 Integration.
In this study five dental models were used. Eight views were captured in order to construct
each model, with the scanner tilted at 45°, thus achieving good detail within the object of in-
terest. This process was conducted by the research group in Perception and Intelligent Control
(PCI). For further details please refer to the Master's final technical report [ 14 ] .
3 Point cloud segmentation
In this section, a brief introduction on three different techniques for segmenting point clouds
is given: RANdom SAmpling and Consensus (RANSAC) [ 15 ] , Region Growing [ 16 ] , and
Maximum-Flow Minimum-Cut (Min-Cut) [ 17 ] .
3.1 RANSAC
The iterative method of RANSAC was proposed by Fischer and Bolles. The technique aims to
estimate the parameters of a mathematical model from a set of observed data using a meth-
od of hypothesis testing. The algorithm is used as a geometric model-based segmentation al-
gorithm, due to its ability to automatically recognize parameterized shapes through the data.