Chapter 24
3D Object Classification, Visual Search
from RGB-D Data
Vadim L. Kondarattsev , Alexander Yu. Kryuchkov ,
and Roman M. Chumak
Abstract In this chapter, we consider the problem of creating a system for processing
3D models obtained using RGB-D sensors for the purpose of semiautomatic selection
and classification of objects and their auto-completion based on visual search. We
have proposed several heuristic preprocessing algorithms for selecting an object
of interest on a scan that contains noise and extraneous objects. To implement the
visual search algorithm, we obtained a modification of the ray casting 3D-shape
feature extraction algorithm. To solve the classification problem, the possibility of
using deep learning architectures based on convolution mechanisms on graphs is
investigated. The information about the object class obtained during the classification
stage is used for faster and more accurate auto-completion. The resulting system has
been tested on real data.
24.1 Introduction
In recent years, space scanning technologies using depth cameras and lidars have
become widespread. This is due to the active development of such areas as
autonomous vehicles, augmented and virtual reality, medical scanning, computer
vision, and robotics. With an increasing number of 3D datasets and various tasks
related to processing such data, creating systems for automatic detection and
auto-completion of objects in 3D scenes becomes especially important.
V. L. Kondarattsev (B)
Moscow Aviation Institute (National Research University), 4, Volokolamskoe shosse, Moscow
125993, Russian Federation
e-mail: vadim@phygitalism.com
V. L. Kondarattsev · A. Yu. Kryuchkov · R. M. Chumak
PHYGITALISM, 2/46 Bol’shaya Sadovaya ul., Moscow 123001, Russian Federation
e-mail: a.kryuchkov@phygitalism.com
R. M. Chumak
e-mail: p4@phygitalism.com
© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2021
L. C. Jain et al. (eds.), Applied Mathematics and Computational Mechanics for Smart
Applications, Smart Innovation, Systems and Technologies 217,
https://doi.org/10.1007/978-981-33-4826-4_24
353
3D Object Classification, Visual Search
from RGB-D Data
Vadim L. Kondarattsev , Alexander Yu. Kryuchkov ,
and Roman M. Chumak
Abstract In this chapter, we consider the problem of creating a system for processing
3D models obtained using RGB-D sensors for the purpose of semiautomatic selection
and classification of objects and their auto-completion based on visual search. We
have proposed several heuristic preprocessing algorithms for selecting an object
of interest on a scan that contains noise and extraneous objects. To implement the
visual search algorithm, we obtained a modification of the ray casting 3D-shape
feature extraction algorithm. To solve the classification problem, the possibility of
using deep learning architectures based on convolution mechanisms on graphs is
investigated. The information about the object class obtained during the classification
stage is used for faster and more accurate auto-completion. The resulting system has
been tested on real data.
24.1 Introduction
In recent years, space scanning technologies using depth cameras and lidars have
become widespread. This is due to the active development of such areas as
autonomous vehicles, augmented and virtual reality, medical scanning, computer
vision, and robotics. With an increasing number of 3D datasets and various tasks
related to processing such data, creating systems for automatic detection and
auto-completion of objects in 3D scenes becomes especially important.
V. L. Kondarattsev (B)
Moscow Aviation Institute (National Research University), 4, Volokolamskoe shosse, Moscow
125993, Russian Federation
e-mail: vadim@phygitalism.com
V. L. Kondarattsev · A. Yu. Kryuchkov · R. M. Chumak
PHYGITALISM, 2/46 Bol’shaya Sadovaya ul., Moscow 123001, Russian Federation
e-mail: a.kryuchkov@phygitalism.com
R. M. Chumak
e-mail: p4@phygitalism.com
© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2021
L. C. Jain et al. (eds.), Applied Mathematics and Computational Mechanics for Smart
Applications, Smart Innovation, Systems and Technologies 217,
https://doi.org/10.1007/978-981-33-4826-4_24
353