Content uploaded by Pablo Gil
Author content
All content in this area was uploaded by Pablo Gil on Sep 09, 2014
Content may be subject to copyright.
A Performance Evaluation of Surface Normals-based Descriptors for
Recognition of Objects Using CAD-Models
C. M. Mateo
1
, P. Gil
2
and F. Torres
2
1
University Institute for Computing Research, University of Alicante, San Vicente del Raspeig, Spain
2
Department of Physics, Systems Engineering and Signal Theory, University of Alicante, San Vicente del Raspeig, Spain
{cm.mateo, pablo.gil, fernando.torres}@ua.es
Keywords:
3D Object Recognition, 3D Surface Descriptors, Surface Normal, Geometric Modelling.
Abstract:
This paper describes astudy and analysis of surface normal-base descriptors for 3D object recognition. Specif-
ically, we evaluate the behaviour of descriptors in the recognition process using virtual models of objects cre-
ated from CAD software. Later, we test them in real scenes using synthetic objects created with a 3D printer
from the virtual models. In both cases, the same virtual models are used on the matching process to find
similarity. The difference between both experiments is in the type of views used in the tests. Our analysis
evaluates three subjects: the effectiveness of 3D descriptors depending on the viewpoint of camera, the ge-
ometry complexity of the model and the runtime used to do the recognition process and the success rate to
recognize a view of object among the models saved in the database.
1 INTRODUCTION
The 3D object recognition process has had important
advances in the last years. In recent works, many ap-
proaches use range sensors to obtain depth of objects
present in a scene. The depth information has per-
mitted to change the techniques and algorithms for
extracting features from image. In addition, this one
has been used to design and createnew descriptors for
identification objects from scene captured by range
sensors (Rusu, 2009) and (Lai, 2013). LIDARSs,
Time of Flight cameras (ToF) or RGBD sensors, such
as Kinect or Asus Xtion PRO Live, provide depth and
allow us to recover the 3D structure of scene from
a single image. The choice of the kind of sensor
is depending on the context and lighting conditions
(indoors, outdoors) and type of specific application
(guided/navigation of robots or vehicles, people de-
tection, human-machine interaction, object recogni-
tion and reconstruction, etc.). Furthermore,the recog-
nition methodology applied to retrieve the 3D object
shape is different depends on whether the object is
rigid or non-rigid. A variety of methods for detec-
tion of rigid and non-rigid objects were presented in
(Wohlkinger et al., 2012) and (Lian et al., 2013), re-
spectively.
In this work, rigid object recognition is done. But
rigid object recognition can be based on visual fea-
tures information such as bounding, skeleton, silhou-
ette, colour, texture, moments, etc. or geometric fea-
tures such as vectors normal, voxels, etc. obtained
from depth information captured from a range sen-
sor. Examples of descriptors for rigid objects based
on geometric features, are: PFH (Point Feature His-
togram) and FPFH (Fast Point Feature Histogram)
(Rusu, 2009); VFH (Viewpoint Feature Histogram)
(Rusu et al., 2010); CVFH (Clustered Viewpoint Fea-
ture Histogram) (Aldoma et al., 2011); and SHOT
(Signature of Histograms of Otientations) (Tombari
et al., 2010). All of them describe the geometry of
an object using normal vectors to its surface which is
represented by a point clouds. Other descriptors such
as ESF (Ensemble of shape Functions) (Wohlkinger
and Vincze, 2011a) and SVDS (shape Distribution
on Voxel Surfaces) (Wohlkinger and Vincze, 2011b);
GRSD (Global Radius based Surface Descriptors)
(Marton et al., 2011) are based on voxels to represent
the object surface. SGURF (Semi-Global Unique Ref-
erence Frames) and OUR-CVFH (Oriented, Unique
and Repeatable CVFH) (Aldoma et al., 2012b) are
also other noteworthy descriptors because they have
the advantage to the ambiguity over the camera roll
angle. SGURF is computed from a single viewpoint
of the object surface and OUR-CVFH is based on a
mix between SGURF and CVFH. CVFH is briefly
discussed below.
In this paper, 3D rigid object recognition based on
object category recognition is done. Also, we have
428
introduced some novelty into the performance shown
in (Wohlkinger et al., 2012) and (Alexandre, 2012).
We have created views from a virtual camera which
captures information of virtual models with different
viewpoints. Afterwards, we have created the 3D rigid
objects from CAD models using 3D printer to test
if the behavioural changes of the descriptors are sig-
nificant. Thereby, the errors in the recognition pro-
cess can be better controlled. Thus, both descriptors,
model and object, are computed from known perfect
geometrical figures. Therefore, the recognition errors
only depend on the geometry of the isolated object
in the scene and the precision of descriptor for mod-
elling and identifying these objects. It is important
emphasize that evaluated descriptors cannot be used
if the scene was not previously segmented and the ob-
jects are localized therein.
The rest of this paper is structured as follows. 3D
descriptors based on geometric information are com-
mented in Section 2. In Section 3, we present the sim-
ilarity measures proposed for associating objects to
models. Experimental results of the descriptors eval-
uation is shown in Section 4 and 5. Finally, section 6,
contains the conclusions.
2 3D DESCRIPTORS
In this paper, we work with isolated rigid objects with
uncluttered backgroundsin indoor scenes. Hence, our
appearancemodel is based on a set of different feature
descriptors. In particular, five descriptors are used in
the experimentation. For each descriptor type, we use
the same training framework. That is the same ob-
jects as dataset or test data. The training framework is
detailed later (Section 4). The descriptors are always
computed over a mesh consists of a point cloud. The
descriptors only include geometric information based
on the surface shape but they do not include colour or
other type of visual features information. The idea is
to evaluate 3D objects recognition methods based on
3D descriptors without using additional appearance
informationsuch as colour and texture from scene im-
age, information position/orientation from geoloca-
tion and odometry techniques obtained. The absence
of colour and texture provides generality for working
with unknown objects and simplifies the runtime in
the recognition task. Frequently, in the industrial en-
vironments are used objects and pieces without this
kind of information. Those are made of metal or plas-
tic material with homogeneous colour and they can
only be differenced by means of geometry and sur-
face features.
The five feature descriptors based on surface nor-
mal vectors: PFH, FPFH, SHOT, VFH and CVFH,
were chosen because they retrieve enough geometri-
cal information of shape. This information will give
us the ability to make further analysis in industrial
pieces. In the literature, descriptors are grouped as
local and global recognition pipeline. The main dif-
ference among these groups is the size of signature
and the number of signatures to describe the surface.
In the first, descriptor is represented by a signature for
each point of surface, but, in the second, it saves all
viewpoint information using one signature for whole
surface. A brief description:
PFH, It is a set of signatures from several local
neighbourhoods. For each point is computed a 3-
tuple, h α, φ, θ i of angles which represent the
relation among normals in their neighbourhood,
according to Darboux frame. Then in order to,
compute each final signature, the method adds the
relations among all points within neighbourhood
in the surface. Therefore the complexity compu-
tational is O
nk
2
. The signature dimensionality
is 125.
FPFH, This is based on the same idea that PFH, it
uses a Darboux frame to make relations among
pair of points within a neighbourhood with ra-
dio r for computing each local surface signature.
This descriptor generates a linear complexity in
the number of neighbours, O(nk). This approxi-
mation changesthe relations among a point and its
neighbours located with a distance smaller than r,
adding a specific weight according to the distance
between point and everyneighbour. The signature
dimensionality is 33.
SHOT, In this descriptor a partitioned spherical grid
is used as local reference frame. For each volume
of the partitioned grid, a signature of the amount
of cosθ
i
between the normal at each point of sur-
face and the normal at the query feature point is
computed. A normalization of descriptor is re-
quired to provide it robustness towards point den-
sity variations. The signature dimensionality is
352.
VFH, It is based on FPFH. Eachsignature consists of
a histogram with two components; one has the an-
gles hα, φ, θi which is calculated as the angular re-
lation between a point’s normal and the normal of
the point cloud’s centroid, and other represent the
angles between the vector determined by the sur-
face centroid and viewpoint. This descriptor has
complexity of O(n). The signature dimensionality
is 308.
CVFH, This descriptor is an extension to VFH. The
basic idea is to identify an object from splitting it
APerformanceEvaluationofSurfaceNormals-basedDescriptorsforRecognitionofObjectsUsingCAD-Models
429
(a) Cone (b) Cube (c) Cylinder (d) Prism (e) Sphere
Figure 1: Primive shapes of the models.
in a set of smooth and continuous regions or clus-
ters. The edges, ridges and other discontinuities in
the surface are not considered because these parts
are more affected by the noise. Thereby, for each
of these clusters is computed its VFH descrip-
tor. CVFH describes a surface as a histogram in
which each histogram item represents the centroid
to surface and the average of the normals among
all points of surface. Again, the dimensionality is
308.
Other descriptors such as Radius-based (RSD and
GRSD) or voxels-based (SVDS and ESF) are not stud-
ied here. This decision was taken because the re-
sults shown in (Aldoma et al., 2012a) and (Alexandre,
2012) that Normal-based descriptors are best with
household object as proving the accumulated recog-
nition rate, ROC curve for recognition and Recall-vs-
(1-Precision).
3 SIMILARITY MEASURES
Similarity measures are used to associate the CAD-
model and the object view. The similarity measures
are defined like distance metrics. Four type of dis-
tance metrics, d
s
= {d
L1
, d
L2
, d
χ
2
, d
H
} are used to
compare the CAD-model, C
j
, which represents a ob-
ject category with the object view in the scene. The
definitions for the four distances are:
d
L1
(p, q) =
n
∑
i=1
p
i
−q
i
(1)
d
L2
(p, q) =
s
n
∑
i=1
(p
i
−q
i
)
2
(2)
d
χ
2
(p, q) =
n
∑
i=1
(p
i
−q
i
)
2
p
i
+ q
i
(3)
d
H
(p, q) =
1
√
2
s
n
∑
i=1
(
√
p
i
−
√
q
i
)
2
(4)
where d
L1
defines the Manhattan distance, d
L2
is
Euclidean distance, d
χ
2
defines Chi-squared distance
and d
H
is Hellinger distance. And n is point dimen-
sions, being p and q two arbitrary points.
Each CAD-model, C
j
is defined by a set of views
C
j
=
c
j1
, c
j2
. . . , c
jr
where r is the number of view-
points from where the CAD-model is observed with
a virtual camera. Furthermore, each view is repre-
sented by a set of descriptors defined as following,
c
jl
=
n
m
jl
1
, m
jl
2
, m
jl
3
, m
jl
4
, m
jl
5
o
where l represents the
view identifier and j the object class defined in the
CAD-model. This set represents a hybrid descrip-
tor composed of five components. A component for
each type of descriptor: PFH, FPHF, SHOT, VFH
and CVFH. Similarly, for each object, O
i
is defined
by a set of views O
i
= {o
i1
, o
i2
, . . . , o
in
}where n is the
number of viewpoints from where the object in scene
is captured using a virtual or real camera. As well,
each view is represented by a set of descriptors de-
fined as following, o
ik
=
v
ik
1
, v
ik
2
, v
ik
3
, v
ik
4
, v
ik
5
where k
represents the view identifier, and i is the object iden-
tifier.
Then, the difference between each component of
the CAD-model descriptor and object descriptor, is
calculated according to equations (1), (2), (3) and (4).
The similarity, d
c
, between object category, C
j
in
the database and the object in scene, is computed by
using the minimum distance for each type of descrip-
tor, following equation (5). The comparison is done
for all models saved in the database.
d
c
(O
i
,C
j
) = min
o
ik
∈O
i
∧ c
jl
∈C
j
n
d
o
ik
, c
jl
o
(5)
d
o
ik
, c
jl
=
q
d
s
(o
ik
, c
jl
)
2
+ d
s
(c
jl
, o
ik
)
2
(6)
where s represents the kind of distance defined in
equation (1), (2), (3) and (4).
4 EXPERIMENTS
Test data were created to analyse the 3D descriptors
behaviour. They were created like a dataset of the 5
basic shapes which are used like models of objects.
They are a sphere, cube, cone, cylinder and triangu-
lar prism (Figure 1). These models represent differ-
ent surfaces without colour, texture or another charac-
teristic different to geometry. Each CAD-model was
ICINCO2014-11thInternationalConferenceonInformaticsinControl,AutomationandRobotics
430
(a) Tesselated-sphere and arbitrary viewpoints
(b) Top, side and another.
Figure 2: (a) Camera poses to obtain views. (b) Virtual and
real objects views from three arbitrary poses, respectively.
created as a point cloud from CAD software. Each
CAD-model represents an object category in order to
recognize. They are represented by a point cloud with
variable number of points, with regards to the view
and the kind of shape.
The correspondence process between model and
object must be consistent. For this reason, in this
paper, we have evaluated this process using CAD-
models. In addition, we did not use keypoints com-
puted from the surface and so the noise due to in-
accuracy in its location is almost eliminated. There-
fore, factors like the repeatable of keypoints with re-
spect to viewpoint variations cannot be produced. We
have used all points in the surface to analyse and
evaluate the descriptors behaviour, thoroughly. If we
had only evaluated the descriptors with a number of
points chosen from surface, i.e. keypoints, the anal-
ysis had been limited to effectiveness of those. The
keypointsmust be chosen to avoidredundantor sparse
information (keypoints close or too far themselves,
respectively). Generally, the descriptors based on
keypoints are efficient but they are little descriptive
and they are not robust to noise. Other descriptors,
such as local/regional or global descriptors are more
suitable to noise. Moreover, they are useful to han-
dle partial/complete surface information and so they
are more descriptive on objects with poor geometric
structure. Therefore, they are more suitable to cate-
gorize objects in a recognition process, as can be seen
here.
In the experiments, geometric transformations are
applied to the point cloud of CAD-models shown in
Figure 1. Geometric transformations simulate view-
point of the objects in scene of real world. Geometric
transformations applied were rotations, translations
and scale changes from different camera poses (Fig-
ure 2). The recognition process consists of a match-
ing process among CAD models and objects in or-
der to associate and identify the object category. The
object category is given by the object greatest simi-
larity between the object and the geometric shape of
a model (Figure 3, Figure 4 and Figure 5), applying
Equation 5.
In order to evaluate the behaviour descriptors and
find which works best in recognition process, we have
planned two type of experiments. Firstly, virtual ob-
jects are created from CAD-models selecting viewsto
build the test database (Figure 3). Thus, at least, we
guarantee that all views created for the test database
are equals to one view of a CAD-model. Secondly,
virtual objects are created from CAD-models apply-
ing one or more transformation on those (Figure 4).
These transformations are chosen to provide different
views to any view used within a model so we ensure
a total difference between test database and models.
In this case, we have worked with 42 and 38 different
views of the test and model database, respectively.
Figure 3 shows a comparison in which the match-
ing process is done combining all descriptors with
all distances for virtual object views without trans-
formations. This comparison allows us to determine
the capacity of similarity measures for classification
of object views in categories according to a CAD-
model. The obtained results report better recognition
when the matching process is done using L1 distances
and the worst results are generated by L2 distance, in
both case is independent from the used 3D descrip-
tor. In addition, L2 distance causes confusion in the
recognition as distance matrices of PFH, FPFH and
SHOT demonstrate.χ
2
and H provide similar results
although H is slightly better.
Figure 4 shows an interesting additional experi-
ment. It consists in reporting recognition results with
regard to the transformation level. The difficulty in
the matching process is increased due to the loss of
similarity among the virtual object views with trans-
formation and the models. In this case, both distance
matrices, VFH and SHOT, report about a growth of
confusion level in the recognition regardless of dis-
tance metric. Furthermore, both PFH and FPFH are
not practically changed their behaviour. Summariz-
ing, CVFH is the most stable descriptor although the
chosen distance metric is different or the object views
are not exactly equal to any model views.
Finally, we have tried out the behaviour of the two
best descriptors using the two best similarity mea-
sures when the recognition process is realized from
real physical objects. In this case, the views for the
test database are obtained by means of acquisition
process from Kinect. In this last experiment, CAD-
models are used to create 5 real physical objects us-
ing a 3D printer. They were created using PLA (PLA:
APerformanceEvaluationofSurfaceNormals-basedDescriptorsforRecognitionofObjectsUsingCAD-Models
431
(a) PFH (L2) (b) PFH (χ
2
) (c) PFH (H) (d) PFH (L1)
(e) FPFH (L2) (f) FPFH (χ
2
) (g) FPFH (H) (h) FPFH (L1)
(i) SHOT (L2) (j) SHOT (χ
2
) (k) SHOT (H) (l) SHOT (L1)
(m) VFH (L2) (n) VFH (χ
2
) (o) VFH (H) (p) VFH (L1)
(q) CVFH (L2) (r) CVFH (χ
2
) (s) CVFH (H) (t) CVFH (L1)
Figure 3: Distance matrix when model set is compared with itself (Model vs Model).
PolyLactic Acid or PolyLActide) filament of 3mm di-
ameter. The print allowed us a precisely controlling
of the size, exact shape and the building material that
objects would have in the scene. This is done be-
cause we would not have an appropriated error han-
dling, if household objects similar to (Rusu, 2009) or
(Alexandre, 2012) had been used in our experiments.
Perhaps, in those cases, the errors in the recognition
process were influenced by the properties of building
material, the capture and digitalized process when the
shapes are not exactly like the CAD-model, etc. For
this reason, we have built our own objects for the test
database. After we have captured from Kinect these
real physical objects using different pose cameras in
the scene. In particular, the test data set has a total of
32 camera views for each object. These viewpoints
represent rotations and translations. The object has
been rotated from 4 different angles (0,
π
6
,
π
3
,
π
2
)rad in
two different axis (in relation of the main axis and
minor axis of the object). In addition, the object has
been translated to 4 different positions which repre-
sent (origin, near, left and right). This way the scale
ICINCO2014-11thInternationalConferenceonInformaticsinControl,AutomationandRobotics
432
(a) PFH (χ
2
) (b) PFH (L1)
(c) FPFH (χ
2
) (d) FPFH (L1)
(e) SHOT (χ
2
) (f) SHOT (L1)
(g) VFH (χ
2
) (h) VFH (L1)
(i) CVFH (χ
2
) (j) CVFH (L1)
Figure 4: Distance matrix when model set is compared with
test set (Model vs Test).
changes have also been considered. The result can be
seen in Figure 5 which shows the matching process
between all objects and all CAD-models.
As the above Figures 4 clearly shown, CVFH is
the most effective to recognize virtual objects. There-
fore, it turns out a good choice to apply it to recog-
nize real physical objects using similar views to those
were registered for the virtual objects as is shown in
Figure 5. A comparison of Figures 4(i)- 4(j) and Fig-
ures 5(c)- 5(d) demonstrate that the presence of varia-
tions, such as present noise, lacking of pointsto define
the surface when the view is captured from camera or
loosing of smoothing surface due to noise points in
the acquisition process, have worsened the matching
(a) VFH (χ
2
) (b) VFH (L1)
(c) CVFH (χ
2
) (d) CVFH (L1)
Figure 5: Distance matrix for matching process among
models and real scenes.
process. Consequently, the distance between a view
and false model are closer to zero. This fact is clearly
observed between cylinder and cone.
5 ANALYSIS AND EVALUATION
OF TIME AND ACCURACY
The recognition process behaviour have been evalu-
ated with regards to the relation between runtime and
accuracy. A complete set of experiments were de-
signed. Summarizing, the recognition process con-
sisted of three steps: a) Building database: Calcula-
tion of descriptors for each view in each model saved
in the database. b) Calculation of descriptors for real
and virtual (test) views. c) Matching between test
views by means of computing difference among all
models views saved in the database and arbitrary test
view.
The runtime of steps a) and b) on the recogni-
tion process is changing and it depends on amount of
points in the view, the number of views per model, the
number of models and the descriptor characteristics.
Thus, we have to measure the runtime cost depend-
ing on detail level of its representation in each point
cloud. Figure 6 shows the runtime for each descrip-
tor depending on the shape. Each graph represents
the runtime of all descriptors for each shape (for each
shape were used 162 views with different amount of
points). On the one hand, as observed, the runtime
dependencywith shape complexityis least-significant
than computational complexity of feature descriptor.
It is because all shapes keep the following relation:
PFH >> FPFH >> SHOT >> CVFH >> VFH.
Although, the shape complexity affects to stability of
local feature descriptors runtime (Figure 6(f)). VFH
APerformanceEvaluationofSurfaceNormals-basedDescriptorsforRecognitionofObjectsUsingCAD-Models
433
300 400 500 600 700 800 900 1000 1100
10
0
10
1
10
2
10
3
10
4
Number of points
Descriptor runtime (msec.)
pfh
fpfh
shot
vfh
cvfh
(a) Cone
300 400 500 600 700 800 900 1000 1100
10
0
10
1
10
2
10
3
10
4
Number of points
Descriptor runtime (msec.)
pfh
fpfh
shot
vfh
cvfh
(b) Cube
300 400 500 600 700 800 900 1000 1100
10
0
10
1
10
2
10
3
10
4
Number of points
Descriptor runtime (msec.)
pfh
fpfh
shot
vfh
cvfh
(c) Cylinder
300 400 500 600 700 800 900 1000 1100
10
0
10
1
10
2
10
3
10
4
Number of points
Descriptor runtime (msec.)
pfh
fpfh
shot
vfh
cvfh
(d) Prism
300 400 500 600 700 800 900 1000 1100
10
0
10
1
10
2
10
3
10
4
Number of points
Descriptor runtime (msec.)
pfh
fpfh
shot
vfh
cvfh
(e) Sphere
(f) Mean and standard errors
Figure 6: Descriptor runtime depending on the shape.
Cone Cube Cylinder Prism Sphere
10
0
10
1
10
2
10
3
10
4
10
5
10
6
Matching runtime(msec.)
shot
pfh
fpfh
cvfh
vfh
(a) Euclidean
Cone Cube Cylinder Prism Sphere
10
0
10
1
10
2
10
3
10
4
10
5
10
6
Matching runtime(msec.)
shot
pfh
fpfh
cvfh
vfh
(b) Chi-squared
Cone Cube Cylinder Prism Sphere
10
0
10
1
10
2
10
3
10
4
10
5
10
6
10
7
Matching runtime(msec.)
shot
pfh
fpfh
cvfh
vfh
(c) Hellinger
Cone Cube Cylinder Prism Sphere
10
0
10
1
10
2
10
3
10
4
10
5
10
6
10
7
Matching runtime(msec.)
shot
pfh
fpfh
cvfh
vfh
(d) Manhattan
Figure 7: Matching runtime for each descriptor depending
on the shape.
and CVFH are the fastest in this comparison.
On the other hand, a study of the balance be-
tween runtime and accuracy is realized in step c).
Firstly, Figure 7 shows the mean runtime in match-
ing process between a test view and models database.
Again, the set of global descriptors (VFH and CVFH)
is faster than others (10
3
times), independently for the
high dimensionality of its signatures. Secondly, Fig-
ure 8 shows the difference between accuracy when
the matching process is made using models such as
test views and when it is made using test views. In
addition, accuracy is less using local descriptors than
global descriptors. Although CVFH has the best ac-
curacy rate, another important issue is the metric se-
shot pfh fpfh cvfh vfh
0
20
40
60
80
100
Accuracy (%)
Model vs Model
Model vs Test
(a) Euclidean
shot pfh fpfh cvfh vfh
0
20
40
60
80
100
Accuracy (%)
Model vs Model
Model vs Test
(b) Chi-squared
shot pfh fpfh cvfh vfh
0
20
40
60
80
100
Accuracy (%)
Model vs Model
Model vs Test
(c) Hellinger
shot pfh fpfh cvfh vfh
0
20
40
60
80
100
Accuracy (%)
Model vs Model
Model vs Test
(d) Manhattan
Figure 8: Accuracy rates for descriptors depending on met-
ric used in matching process.
lection. In terms of runtime, this selection is not out-
standing (Figure 7), but it is important in terms of
accuracy (Figure 8). In the experiments, model vs
model represented in Figure 3, a 20% increase of ac-
curacy rate is obtained. When L1 is used as observed
in Figure 8(a) - 8(d). Nevertheless, the best result is
obtained using χ
2
in th experiment, model vs test rep-
resented in Figure 4. In this case, a 5% increase of
accuracy is achieved.
ICINCO2014-11thInternationalConferenceonInformaticsinControl,AutomationandRobotics
434
6 CONCLUSIONS
This paper discusses the effectiveness of using 3D de-
scriptors based on normals to surfaces in order to rec-
ognize geometric objects. 3D descriptors were used
for real physical and virtual objects recognition by
means of matching with virtual geometric models. A
total of 6028 tests have been done. Where 3800 tests
(4 different distances, 5 descriptors, 5 shapes and 38
views per shape) are from the model-vs-model ex-
periment, 2100 tests (2 different distances, 5 descrip-
tors, 5 shapes and 42 views per shape) are from the
model-vs-test experiment and 128 tests (2 different
distances, 2 descriptors, one shape and 32 views) are
from the model-vs-real-physical-object experiment.
SHOT and FPFH are run in CPU-based parallel im-
plementation. The computer specification is Intel
Core i7-4770k processor, equipped with 16GB of sys-
tem memory and GPU is Nvidia GeForce 770GTX.
The effectiveness of recognition process is evaluated
by measuring the runtime and the precision to achieve
success rate of the recognition process. Those are de-
pending on the type of descriptor, resolution of the
point cloud which represents each object, and the
level of accuracy required for the recognition.
ACKNOWLEDGEMENTS
The research leading to these result has received
funding from the Spanish Government and European
FEDER funds (DPI2012-32390)and the Valencia Re-
gional Government (PROMETEO/2013/085).
REFERENCES
Aldoma, A., Marton, Z.-C., Tombari, F., Wohlkinger, W.,
Potthast, C., Zeisl, B., Rusu, R. B., Gedikli, S., and
Vincze, M. (2012a). Tutorial: Point cloud library:
Three-dimensional object recognition and 6 dof pose
estimation. In IEEE Robot. Automat. Mag., vol-
ume 19, pages 80–91.
Aldoma, A., Tombari, F., Rusu, R. B., and Vincze,
M. (2012b). Our-cvfh - oriented, unique and re-
peatable clustered viewpoint feature histogram for
object recognition and 6dof pose estimation. In
DAGM/OAGM Symposium, pages 113–122.
Aldoma, A., Vincze, M., Blodow, N., Gossow, D., Gedikli,
S., Rusu, R.B., and Bradski, G. R. (2011). Cad-model
recognition and 6dof pose estimation using 3d cues.
In IEEE International Conference on Computer Vision
Workshops, ICCV 2011 Workshops, Barcelona, Spain,
November 6-13, 2011, pages 585–592.
Alexandre, L. A. (2012). 3D descriptors for object and cate-
gory recognition: a comparative evaluation. In Work-
shop on Color-Depth Camera Fusion in Robotics at
the IEEE/RSJ International Conference on Intelligent
Robots and Systems (IROS), Vilamoura, Portugal.
Lai, K. (2013). Object Recognition and Semantic Scene
Labeling for RGB-D Data. PhD thesis, University of
Washington, Washington, USA.
Lian, Z., Godil, A., Bustos, B., Daoudi, M., Hermans, J.,
Kawamura, S., Kurita, Y., Lavou´e, G., Van Nguyen,
H., Ohbuchi, R., Ohkita, Y., Ohishi, Y., Porikli, F.,
Reuter, M., Sipiran, I., Smeets, D., Suetens, P., Tabia,
H., and Vandermeulen, D. (2013). A comparison
of methods for non-rigid 3d shape retrieval. Pattern
Recogn., 46(1):449–461.
Marton, Z.-C., Pangercic, D., Blodow, N., and Beetz, M.
(2011). Combined 2d-3d categorization and clas-
sification for multimodal perception systems. I. J.
Robotic Res., 30(11):1378–1402.
Rusu, R. B. (2009). Semantic 3D object maps for every-
day manipulation in human living environments. PhD
thesis, Technical University Munich.
Rusu, R. B., Bradski, G.R., Thibaux, R., and Hsu, J. (2010).
Fast 3d recognition and pose using the viewpoint fea-
ture histogram. In IROS, pages 2155–2162. IEEE.
Tombari, F., Salti, S., and Stefano, L. D. (2010). Unique
signatures of histograms for local surface descrip-
tion. In Proceedings of the 11th European Conference
on Computer Vision Conference on Computer Vision:
Part III, ECCV’10, pages 356–369, Berlin, Heidel-
berg. Springer-Verlag.
Wohlkinger, W., Aldoma, A., Rusu, R. B., and Vincze, M.
(2012). 3dnet: Large-scale object class recognition
from cad models. In ICRA, pages 5384–5391. IEEE.
Wohlkinger, W. and Vincze, M. (2011a). Ensemble of
shape functions for 3d object classification. In RO-
BIO, pages 2987–2992. IEEE.
Wohlkinger, W. and Vincze, M. (2011b). Shape distri-
butions on voxel surfaces for 3d object classification
from depth images. In ICSIPA, pages 115–120. IEEE.
APerformanceEvaluationofSurfaceNormals-basedDescriptorsforRecognitionofObjectsUsingCAD-Models
435
