24 3D Object Classification, Visual Search …
359
of difference is defined for any two models:
d = ρ
M p,s , M k,l
∈ [0; 1].
The difference measure is based only on the model geometry. The more dissimilar
the models are in shape, the greater the d value. Our goal is to find the set of models
R r
M p,s , O
⊂ O, r = 1, 2, . . . , n, where elements of the set are the most similar
models for the source model M p,s . We will construct this set in the following way:
R r
M p,s , O
=
k
i=1
R
i
,
R
i
= arg
min
M∈O/
i−1
l=1 R l
ρ
M p,s , M
,
0
i=1
R
i
= ∅,
k = min
k∈{1,2,...,n}
k:
k
i=1
R
i
≥ r
.
If the number of models in union
k
i=1 R
i is less than r , then we will randomly
choose r −
k
i=1 R
i
models from the set R
k .
24.4 Data Preprocessing
In this section, preprocessing of polygonal models is considered in Sect. 24.4.1,
while point cloud classification is described in Sect. 24.4.2.
24.4.1 Preprocessing of Polygonal Models
For more clarity and tests on real data, we scanned a small area of the room using
a depth camera. The result of scanning in the form of a polygon model is shown in
Fig. 24.2.
We assumed that the object of interest (table) is located near of bounding boxes
geometric center. Here, under the bounding box, we understand AABB of the entire
scan.
If we look at the source data, we can notice several things that prevent us from
selecting the desired object:
359
of difference is defined for any two models:
d = ρ
M p,s , M k,l
∈ [0; 1].
The difference measure is based only on the model geometry. The more dissimilar
the models are in shape, the greater the d value. Our goal is to find the set of models
R r
M p,s , O
⊂ O, r = 1, 2, . . . , n, where elements of the set are the most similar
models for the source model M p,s . We will construct this set in the following way:
R r
M p,s , O
=
k
i=1
R
i
,
R
i
= arg
min
M∈O/
i−1
l=1 R l
ρ
M p,s , M
,
0
i=1
R
i
= ∅,
k = min
k∈{1,2,...,n}
k:
k
i=1
R
i
≥ r
.
If the number of models in union
k
i=1 R
i is less than r , then we will randomly
choose r −
k
i=1 R
i
models from the set R
k .
24.4 Data Preprocessing
In this section, preprocessing of polygonal models is considered in Sect. 24.4.1,
while point cloud classification is described in Sect. 24.4.2.
24.4.1 Preprocessing of Polygonal Models
For more clarity and tests on real data, we scanned a small area of the room using
a depth camera. The result of scanning in the form of a polygon model is shown in
Fig. 24.2.
We assumed that the object of interest (table) is located near of bounding boxes
geometric center. Here, under the bounding box, we understand AABB of the entire
scan.
If we look at the source data, we can notice several things that prevent us from
selecting the desired object: