17-03-2014, 10:29 PM
Abstract
In this paper we present significant new work on a vision system
designed to recognize 3D objects in a depth map. As previously
described [20], the system was capable of recognizing parametric
surfaces of various types. We have added the ability to find
parametric surface intersection curves and use these disparate
types of information to index into an object database.
From a depth map containing one or more objects, local surface
and discontinuity estimates are determined. These are used in a
series of layered and concurrent parameter space transforms to
extract the global features present in the image. The final transform
is a mapping into an object database. An iterative refinement
technique, motivated by work in connectionist systems, is
used to integrate the evidence at each level. Thus fundamentally
different types of evidence can be simultaneously extracted, then
integrated to form a consistent interpretation of an image.
r o d u c t ion
A difficult problem in computer vision is that of recognizing a
part in a bin or other complex environment and then estimating
its location and orientation so that a robot hand can clasp the part
and manipulate it. The 3D object recognition problem ranges in
difficulty from a scene that contains one, possibly self-occluding,
object, to the most involved problem, a scene with overlapping
parts, where the parts are self-occluding and occluding one
another, i.e., the above mentioned bin-picking problem. Apart
from the configuration of the objects in the input, the difficulty
of the recognition problem is determined by the size of the set of
possible objects. Our goal is to build a vision system capable of
recognizing in a large domain of objects. Eventually, this should
be accomplished in such a hostile environment as a bin of parts.
Human beings are able to process information contained in a
complex scene effectively and seemingly without effort focus on
those clues or features that are pertinent to the task at hand. Also,
they are very efficient at combining these local clues to arrive at
global interpretations. Using a machine to recognize objects, it
is hard to determine apriori which clues are important and which
are not. We propose to extract an abundance of features of
different type from different sources of input; this information
can be fused to determine the content of the 3D scene, that is, to
obtain a consistent explanation of the data. Examples of different
input sources are range, intensity images, and possibly tactile
data. The different types of information that can be extracted
from these inputs include surface parameters, surface-intersection
curves, color, surface markings, etcetera.
The organization of the vision system is illustrated in Figure 1, it
is directed at supporting and even encouraging evidence integration.
Supporting and conflicting feature hypotheses can interact
at various stages of recognition in an effort to reach a consensus.
The current incarnation of the system uses only one input source,
a depth map. From this map the salient primitive features are
reconstructed, i.e., we extract the parameters of object surfaces,
surface-intersection curves, and object limbs. These fundamentally
different types of information are combined to recognize 3D
objects. Hence, a solid beginning to the general system depicted
in Figure 1 is presented.
In this paper we present significant new work on a vision system
designed to recognize 3D objects in a depth map. As previously
described [20], the system was capable of recognizing parametric
surfaces of various types. We have added the ability to find
parametric surface intersection curves and use these disparate
types of information to index into an object database.
From a depth map containing one or more objects, local surface
and discontinuity estimates are determined. These are used in a
series of layered and concurrent parameter space transforms to
extract the global features present in the image. The final transform
is a mapping into an object database. An iterative refinement
technique, motivated by work in connectionist systems, is
used to integrate the evidence at each level. Thus fundamentally
different types of evidence can be simultaneously extracted, then
integrated to form a consistent interpretation of an image.
r o d u c t ion
A difficult problem in computer vision is that of recognizing a
part in a bin or other complex environment and then estimating
its location and orientation so that a robot hand can clasp the part
and manipulate it. The 3D object recognition problem ranges in
difficulty from a scene that contains one, possibly self-occluding,
object, to the most involved problem, a scene with overlapping
parts, where the parts are self-occluding and occluding one
another, i.e., the above mentioned bin-picking problem. Apart
from the configuration of the objects in the input, the difficulty
of the recognition problem is determined by the size of the set of
possible objects. Our goal is to build a vision system capable of
recognizing in a large domain of objects. Eventually, this should
be accomplished in such a hostile environment as a bin of parts.
Human beings are able to process information contained in a
complex scene effectively and seemingly without effort focus on
those clues or features that are pertinent to the task at hand. Also,
they are very efficient at combining these local clues to arrive at
global interpretations. Using a machine to recognize objects, it
is hard to determine apriori which clues are important and which
are not. We propose to extract an abundance of features of
different type from different sources of input; this information
can be fused to determine the content of the 3D scene, that is, to
obtain a consistent explanation of the data. Examples of different
input sources are range, intensity images, and possibly tactile
data. The different types of information that can be extracted
from these inputs include surface parameters, surface-intersection
curves, color, surface markings, etcetera.
The organization of the vision system is illustrated in Figure 1, it
is directed at supporting and even encouraging evidence integration.
Supporting and conflicting feature hypotheses can interact
at various stages of recognition in an effort to reach a consensus.
The current incarnation of the system uses only one input source,
a depth map. From this map the salient primitive features are
reconstructed, i.e., we extract the parameters of object surfaces,
surface-intersection curves, and object limbs. These fundamentally
different types of information are combined to recognize 3D
objects. Hence, a solid beginning to the general system depicted
in Figure 1 is presented.