24 3D Object Classification, Visual Search …
355
However, in recent years, with the rapid development of the field of deep learning,
approaches based on the use of deep neural networks, which allow combining the
stage of feature selection and classification into one algorithm, are becoming more
popular. One of the ways to build taxonomy of deep architectures for processing
a point cloud can be a method based on the division of architectures into classes
depending on how the input data is processed. Thus, for example, in [4], authors
divide all architectures into direct methods and indirect ones.
Indirect architectures do not process the initial point cloud, but some intermediate representation (it can be voxelized models or a set of point cloud images in
RGB-D format). Most often, indirect methods are inferior in quality to direct ones,
and besides, they are more expensive in terms of memory (you need memory to
store intermediate forms of data) and of time (you need time to get an intermediate
representation).
On other hand, the direct methods of deep learning for solving the problems of
processing a point cloud, similar to the study [4], can be divided into different classes
of methods, depending on the types of operators—hidden layers used in building of
the architecture or depending on the modification of the basic architecture of deep
learning, based on which specific methods have been developed.
In order to select a deep learning model for solving the point cloud classification problem, we performed a comparative analysis of quality metrics from various
sources mentioned in [4]. The comparison results are shown in Table 24.1. The
comparison was made for datasets [5–8]. The following metrics from [4] were
considered for comparing models:
• Overall accuracy.
• Mean accuracy.
• Mean intersection over union.
As a result, the LDGCNN model [1] was chosen for the practical implementation
in the automatic object classification system. This method is a direct method, and it
is based on modification of the GCNN architecture [9].
24.3 Formal Statement of the Problem
Our global task is divided into three stages:
1. Preprocessing of the polygon models.
2. Classification of the selected polygon model.
3. Autocomplete based on a search among similar models of the same class that
was defined at the previous stage.
Before proceeding to the description of the main stages, it is necessary to describe
how this data is presented in a more formal form. The classification of polygonal
models will be solved as the classification of the point cloud problem. At the moment,
this solution is associated with the possibility of using a broader class of algorithms
355
However, in recent years, with the rapid development of the field of deep learning,
approaches based on the use of deep neural networks, which allow combining the
stage of feature selection and classification into one algorithm, are becoming more
popular. One of the ways to build taxonomy of deep architectures for processing
a point cloud can be a method based on the division of architectures into classes
depending on how the input data is processed. Thus, for example, in [4], authors
divide all architectures into direct methods and indirect ones.
Indirect architectures do not process the initial point cloud, but some intermediate representation (it can be voxelized models or a set of point cloud images in
RGB-D format). Most often, indirect methods are inferior in quality to direct ones,
and besides, they are more expensive in terms of memory (you need memory to
store intermediate forms of data) and of time (you need time to get an intermediate
representation).
On other hand, the direct methods of deep learning for solving the problems of
processing a point cloud, similar to the study [4], can be divided into different classes
of methods, depending on the types of operators—hidden layers used in building of
the architecture or depending on the modification of the basic architecture of deep
learning, based on which specific methods have been developed.
In order to select a deep learning model for solving the point cloud classification problem, we performed a comparative analysis of quality metrics from various
sources mentioned in [4]. The comparison results are shown in Table 24.1. The
comparison was made for datasets [5–8]. The following metrics from [4] were
considered for comparing models:
• Overall accuracy.
• Mean accuracy.
• Mean intersection over union.
As a result, the LDGCNN model [1] was chosen for the practical implementation
in the automatic object classification system. This method is a direct method, and it
is based on modification of the GCNN architecture [9].
24.3 Formal Statement of the Problem
Our global task is divided into three stages:
1. Preprocessing of the polygon models.
2. Classification of the selected polygon model.
3. Autocomplete based on a search among similar models of the same class that
was defined at the previous stage.
Before proceeding to the description of the main stages, it is necessary to describe
how this data is presented in a more formal form. The classification of polygonal
models will be solved as the classification of the point cloud problem. At the moment,
this solution is associated with the possibility of using a broader class of algorithms