Image Processing Reference
3.4 Feature Sampling Using NARF
Semi-automated 3D point cloud segmentation results are highly dependent upon parameter
set-up. Whenever possible, it is the user's responsibility to provide proper values for such
parameter set as part of the algorithm's prior. In some algorithms, those parameters are points
which are representative for the regions contained in the point set, and selecting them can be
a demanding task.
In order to overcome such limitation, an automatic or semi-automatic point selection tech-
nique can be applied.
the necessity of generating a structured subsampling of the data, taking into account elements
such as spatial density (including empty areas) and robustness to noise.
According to the authors, in order to obtain NARF descriptors one must:
• calculate a normal aligned range value patch in the point, which is a small range image
with the observer looking at the point along the normal,
• overlay a star patern onto this patch, where each beam corresponds to a value in the inal
descriptor, that captures how much the pixels under the beam change,
• extract a unique orientation from the descriptor,
• and shift the descriptor according to this value to make it invariant to the rotation.
Resulting points are not overcrowded in high variation areas, but retain a well-defined rep-
resentation of the information contained in the original cloud. Additionally, each feature vec-
tor contains the mean orientation of the local patch that it represents.
3.5 The Hybrid Technique
This method implemented is a contribution for semi-automatic segmentation of point cloud
based in Min-Cut, NARF, and Region Growing.
Min-Cut is an efficient and powerful technique for point cloud segmentation; however, this
method has a number of disadvantages as it is the prior knowledge of the center of object to be
segmented and the radius value. Due to these drawbacks, we propose a method to achieve a
semi-automatic segmentation without prior knowledge of the location of the objects, supply-
ing part of that information by the means of NARF.
On the other hand, Region Growing segmentation based in local surface orientation and
curvature is an efficient technique for segmenting point clouds of dental models, efectively
separating the gum and teeth.
In our first work, a hybrid technique was constructed using NARF and Min-Cut, aimed to
work. The streamline of the technique can be summarized as shown below:
FIGURE 4 This is an example segmentation in outdoor environment using the first work with
the hybrid technique (for easier visualization background is subsample).