You are on page 1of 10

Augmented Reeb Graphs for Content-based Retrieval of 3D Mesh Models

Tony Tung and Francis Schmitt


Signal and Image Processing Department, CNRS UMR 5141
Ecole Nationale Supérieure des Télécommunications
46, rue Barrault, 75634 Paris Cedex 13, France
{tony.tung, francis.schmitt}@enst.fr

Abstract quisition sequences of the object, are not well adapted to 3D


queries. The approaches such as curvatures histograms [8]
This article presents an improved method of 3D mesh mod- or EGI (Extended Gaussian Image) and its variants [9, 10]
els indexing for content-based retrieval in database with describe too locally the geometry of the objects. As a matter
shape similarity and appearance queries. The approach of fact they do not keep any topological information (mesh
is based on the Multiresolutional Reeb Graph matching connectivity) and are highly sensitive to mesh modifications
presented in [1]. The original method only takes into ac- (noise on vertices position, decimation, etc...). In the other
count topological information what is often not sufficient hand, global approaches such as cords histograms [5] or
for effective matchings. Therefore we proposed to augment shape distributions [6] works quite well for the matching
this graph with geometrical attributes. We also provide a of primitive shapes but produce a too coarse description of
new topological coherence condition to improve the graph the 3D objects shape to be a sharp retrieval tool when com-
matching. Moreover 2D appearance attributes and 3D fea- paring similar complex shapes. Similar comments can be
tures are extracted and merged to improve the estimation done for the Fourier based descriptors and moments cal-
of the similarity between models. Besides, all these new culations [3, 7]. To obtain an intuitive description of the
attributes are user-dependent as they can be weighted by shapes, the use of a skeletal or graph representation is very
variable terms. We obtain a flexible multiresolutional and attractive. These approaches permit us to keep the topology
multicriteria representation called Augmented Reeb Graph of the objects which is a rich information for matching pur-
(ARG). Good preliminary results have been obtained in pose. Unfortunately, difficulties lies in the construction of
a shape-based matching framework compared to existing the skeleton which is either time consuming, too sensitive
methods based only on statistical measures. In addition, to noise on the shape surface, or need a source point which
our study lead us to an innovative part matching scheme favour a direction. Besides, most of the time no matching
based on the same approach as our Augmented Reeb Graph scheme is provided [11, 12, 13], or no multiresolutional ap-
matching. proach [14, 15, 16] whom the properties are useful for ac-
curate similarity calculations and for various applications as
presented in this article. The Multiresolutional Reeb Graph
1. Introduction (MRG) [1] appears to be a very promising starting point of
research, in particular because it is an efficient mean to ex-
The development of new 3D applications has become ploit different techniques and merge the features as we will
easier and tend to generalise due to the permanent progress show in the next sections. The Reeb Graph of a 3D object
of numerical technologies. Database of 3D objects has is represented by a unidimensional skeleton. It is built us-
begun to proliferate for multimedia applications [23, 24, ing a function µ defined over the object surface considered
25, 26] as well as for more industrial ones such as in as a compact manifold, and therefore relies on the mesh
CAD (e.g. motorcar companies). 3D objects indexing and connectivity. The graph is robust in case of mesh modifica-
content-based search appear as a natural answer for those tion thanks to the assumed integral function over the whole
who want to navigate in such huge database. As a conse- object [1]. The surface of the object is divided in regions
quence, queries have to be adapted to this new kind of data. according to value intervals of µ, and a node is associated
Scientific literature only presents few methods dealing with to each connected component of the regions. The graph
3D shape similarity queries. The techniques derived from structure is then obtained by linking the nodes of the con-
the 2D, such as colour histogram or texture distribution nected regions. Afterwards the MRG is constructed hier-
taken from texture maps or from colour images of the ac- archically from the finest level of resolution to the coarsest
level. And similarity calculation between MRG are done is easy to calculate and stable to small noise. In the other
from the coarsest level to the finest level. Keeping the ad- hand, this function provides a Reeb Graph which is not flex-
vantages of the multiresolutional representation offers by ible enough for our study, as explained next. The chosen
the MRG, we have provided topological, geometrical and
visual (colour and texture) informations to the node graphs
to obtain a hierarchical description of the models, which is
both global and local and is called Augmented Reeb Graph
(ARG). Therefore the 3D objects indexing in database will
consist on coding in feature vectors the information associ-
ated to the nodes of the Reeb graphs. Tests of shape based
retrieval were performed upon a database of high resolution
3D models [22, 23] and preliminary experiments were per-
formed concerning the application of part matching.
The MRG construction is presented in the next section.
The graph matching procedure and the choice of the nodes
attributes are shown in Section 3. Section 4 deals with
the similarity calculation. The part matching scheme is
presented in Section 5, and the results performed on our Figure 1. Example of µ functions: height func-
database are described in Section 6. We conclude in Sec- tion (left), distance to centre of mass (middle),
tion 7 with a discussion about the possible improvement of geodesic distance integral (right).
the methods.

function µ in [1] has rotation invariance property and is de-


2. Multiresolutional Reeb Graph overview
fined as the sum of the geodesic distance g(v, p) from v to
the others points {p|p ∈ S}:
According to the Morse theory, a continuous function
defined on a closed surface characterise the topology of the
Z
surface on its critical points [2]. Therefore, a Reeb graph µ(v) = g(v, p)dS.
p∈S
can be obtained assuming a function µ calculated over the
3D object surface. The multiresolutional aspect results from The authors [1] chose to normalise the function with
the dichotomic discretization of the function values and maxp∈S µ(p) for the denominator. To exploit the full dy-
from the hierarchical collection of Reeb graphs defined at namic of µ(v) over [0, 1], we have normalised the function
each resolution. with its min and max values:

2.1. The function µ µ(v) − minp∈S µ(p)


µN (v) = .
maxp∈S µ(p) − minp∈S µ(p)
Our study concerns objects which are 3D compact man-
ifold meshes of triangles with vertices located in a Carte- The Reeb graph is obtained by partitioning the object sur-
sian frame R(x, y, z). The authors of [15] present different face following regular intervals of µ values, and linking the
continuous functions which can be used for the Reeb graph connected regions to each other. In each interval, a node
construction (cf. Figure 1). Assuming a point v(x, y, z) on is associated to a set of connected triangles (cf. Figure 2).
the surface S of an object, the height function µ defined Based on geodesic distance calculation, the resulting Reeb
as µ(v(x, y, z)) = z is well adapted for elongated stand graph of a human model with arms lifted up in the air and
models as human representations. This function is easy to a human model with arms along the body will look similar
compute but by definition has a strong orientation depen- as maximal values of µ stay at the extremity of the arms.
dence. Therefore it is not suited for objects for which the On the other hand, the function µ defined as the distance to
points dispersal is perpendicular to the axis z. The function the centre of mass will lead to a different graph. As a mat-
µ defined as the geodesic distance from curvature extrema ter of fact, it can be stated that the choice of the function µ
is based on growing regions associated to local Gaussian depends on the user’s queries.
curvatures on seed vertices. The results highly depend on 2.2. Multiresolutional aspect
the position of the seeds and require the local curvature cal-
culation which is not obvious on every objects, especially A graph of double or half resolution is respectively ob-
those coming from the Internet. The function defined as the tained by the division or the fusion of nodes [1]. A parent
distance of a surface point v to the centre of mass G of the node generates n children nodes with n ≥ 1 at a higher
object µ(v) = d(G, v), where d is the Euclidean distance, level of resolution (cf. Figure 3).
• the parents of m and n have been matched at the pre-
vious level of resolution

• m and n are in the same interval of the function µN

• m and n have the same labels, meaning they both be-


long to branches which have been matched to each
other. In deed, when two nodes are matched, a same
label is propagated along the branches they belong
to. Therefore neighbouring nodes belonging to a same
branch have the same labels. In this way, we ensure
that a branch is entirely matched to a branch [1].

It is also assumed that a node can be matched to a group


Figure 2. Reeb graph: 3D model (left), function µN
of topologically consistent nodes [1]. However the simple
(middle), Reeb graph; edges in green; nodes in
statement of these criteria is not sufficient to ensure that two
red (right) (model from [23]).
branches will still be matched together when the resolution
gets higher (cf. Figure 4). As a consequence, we add a
new topological consistency criterion: if the parents of the
neighbours of m and n have been matched, then they must
have been matched together.

Figure 3. MRG construction by merging µN inter-


vals. Nodes from a level of resolution r merge in
parent nodes at a higher level of resolution (from
left to right: r = 2, 1 and 0).
Figure 4. Matching nodes (left). At a higher level
of resolution, branches mixing can happened. To
avoid this, the matching of the parents of the
3. Augmented Reeb Graph matching neighbours have to be tested (right).

The matching strategy of two MRGs is based on the


detection of nodes with similar topological configurations
in term of connections to their neighbours nodes and par- 3.2. Geometrical attributes
ents/children relations. The aim is to maximise the number
From a topological point of view, the matching of the
of corresponding nodes. For this purpose a set of topolog-
graphs could be acceptable. But in our framework these
ical consistency criteria were stated. The pairs of nodes
matching are not well adapted, because topologically speak-
which verify the criteria are candidates for the matching.
ing an arm worth a leg (they are limbs linked to the body),
And a loss function determines the best matching nodes
whereas we would like to get more accuracy in the match-
among the candidates. To improve the matching of the orig-
ing to efficiently compare complex 3D models.
inal MRG method, we have introduced an additive topolog-
The topological consistency criteria are such defined that if
ical consistency criteria and added geometrical attributes to
two nodes are not matched together at the coarsest level of
the nodes. Finally a function estimates the similarity be-
resolution, whereas they should be, then there is no chance
tween the two Augmented Reeb Graphs.
that their children nodes will be matched together at a finer
3.1. Addititional topological consistency resolution. Therefore it is crucial that nodes are matched the
best they can from the beginning of the consistency tests.
We consider the graph comparison of two objects M and For this reason, we propose to provide localisation infor-
N . Topological consistency criteria permit to find the pairs mation to the nodes. A node m is associated to a set of
of nodes (m, n), with m ∈ MRG(M ) and n ∈ MRG(N ), connected triangles, so it can be located at their barycen-
which could possibly be matched afterwards in the proce- tric position (cf. Figures 2 and 3). We choose a spherical
dure. The nodes m and n are topologically consistent if: coordinate representation (r, θ, ϕ):
• r(m) ≥ 0 is the distance to the node at the centre of possibility to choose either to match models having exactly
the frame, the same shapes, or to let more freedom by allowing the
matching between approximative similar shapes (typically
• θ(m) = arctan xy ∈ [− π2 , π2 ],

two human models in different positions). And this can be
done by simply adapting the weighting terms to the desired
• ϕ(m) = arccos zr ∈ [0, π].

queries. Besides, it is also possible to replace the absolute
Using the example of a human model (cf. Figure 5) whom values used to calculate the differences by quadratic dis-
the axis x would get from the bottom to the top of the body, tances, or any other dedicated functions to emphasise the
axis y oriented from left to right and axis z from behind to differences or similarities between two nodes.
front, we can observe that θ allows to identify the left from Figure 6 shows the gain obtained in the matching of two
the right, ϕ distinguish arms from legs and r could be de- graphs by introducing the localisation information in the
cisive in a ambiguous configuration (an arm up in the air nodes.
could be taken as an head). It becomes necessary that ob-

Figure 6. Left: without geometrical information,


legs can be matched with arms as there are topo-
Figure 5. Nodes located in spherical coordinates logically equivalent (blue links denote matching
(model from [23]). nodes). Right: by adding geometrical informa-
tion, arms and legs are well matched (models from
[23]).
jects are oriented towards the principal directions (x, y, z).
For this purpose, we decide either to use the natural axis of
the object centred to the centre of mass with a coordinates
normalisation, or to apply a classical spatial alignment tech- 4. Similarity calculation
nique such as the Principal Component Analysis to turn the
object towards the principal axis [3]. Assuming known the list of matching nodes, a function
is required to measure the similarity between two 3D mod-
3.3. The loss function
els. First, a function calculates the similarity between each
The loss function loss(m, n) quantifies the difference matched nodes. The original function introduced by [1]
between two nodes m and n. It is applied on every groups has been enriched with various features extracted from the
of topologically consistent nodes and returns those which mesh models. Using the same approach as in Section 3.2
will be effectively matched by minimising the function: for the geometrical attributes, these 2D/3D extracted fea-
tures are weighted to allow different types of queries and to
loss(m, n) = αl · |a(m) − a(n)| + βl · |l(m) − l(n)|+ significantly refine the similarity calculus between two ob-
γl · |r(m) − r(n)| + δl · |θ(m) − θ(n)| + l · |ϕ(m) − ϕ(n)|. jects. This results in the Augmented Reeb Graph represen-
tation. The next sections present the proposed nodes simi-
a is the relative area of the set of triangles associated to a
larity function, the introduced 2D/3D object features (rela-
node. l stands for the complexity of a node and is linked
tive volume, cords lengths, curvatures and colour) and the
to the number of its children as described in [1]. αl , βl , γl ,
resulting graphs similarity function.
δl and l are weighting terms which influence the attributes
according to the type of desired queries. This aspect gives 4.1. Similarity of nodes
lots of freedom to the matching of nodes. Associated to the
chosen function µ defined as the integral of geodesic dis- The calculation of the similarity function between two
tance, the flexibility of the method can be illustrated by the nodes m and n is defined as in [1] :
sim(m, n) = αs ·min(a(m), a(n))+βs ·min(l(m), l(n)).
To improve the retrieval of similar objects with more flex-
ible queries and a larger set of global and local character-
istics we introduce the following geometric and visual fea-
tures extracted from the section of the object associated to
each node m and at each level of resolution:
Figure 7. Volume associated to a node obtained
• v(m), the relative volume associated to m,
by summing the volume of the three types of tetra-
• cords(m), a statistic measure of the extent of the sur- hedra included in the section.
face region associated to m,
• curvm (m) and curvg (m), a statistic of the mean and
Gaussian curvatures estimated on the triangulated sur- 4.3. Cords measurement
face associated to the node m, The cords measure introduced in [5] is represented by a
• colorR,G,B (m), a statistic of the texture/colour associ- normalised histogram of the distance from the triangles cen-
ated to the node. tres to the centre of mass of the object. Applied to a node n
at the finest level of resolution, the histogram cord(n) is the
Features are extracted at the finest level of resolution and as- histogram of the distances between the triangle centres as-
sociated to the corresponding nodes. The multiresolutional sociated to the node n, and the position of their barycenter.
aspect of the measures is ensured by the hierarchical rela- Each contribution to the histogram is weighted by the trian-
tion between parents and children nodes. Our idea is to pro- gle area associated to the cord. Therefore, the histogram is
vide to the nodes all of the relevant characterisations that normalised by the factor R1 · area(S)1
where area(S) is the
could be calculated on a set of triangles. Our choices are total surface area of the object. The histogram cord(m) of
not exhaustive but reflect the current frequently used tech- a parent node m at coarser resolution is thus the histogram
niques. Moreover, it is possible to favour one or several of the cords to the barycenter of the node m obtained using
measures by adaptating the weighting terms. This again, let the barycentres of the children of m. At the coarsest level
a lot of freedom to the users for their queries. of resolution r = 0, the histogram cord(m) is equivalent
4.2. Volume calculation to the approach in [5] applied on the entire mesh. This sta-
tistical measure gives the information on the extent of a set
Let’s consider the interval [µ1 , µ2 ] of the function µ as- of triangles. Hence it is a signature of the mesh geometry
sociated to the node m. The relative volume v(m) is defined associated to a node. The use of statistical methods based
as: on histograms to describe an object, as the cords [5] or the
1 vol(m) shape distribution [6], is well adapted for simple shapes.
v(m) = · ,
R vol(S) In our context and from a certain level of resolution, nodes
where vol(S) represents the total volume of the model, and are mainly locally associated to shapes such as cylinders or
vol(m) stands for the volume of the section of m. vol(S) is paraboloids which justify the relevant use of these geomet-
the sum of the tetrahedra formed by the oriented triangles of rical measures.
the mesh and the centre of mass of the object [4]. vol(m) is
4.4. Curvatures measurement
obtained by summing 1) the signed volume of the tetrahedra
formed by the triangles associated to m and the barycenter Local geometric properties of the surface are represented
of these triangles, 2) the signed volume of the tetrahedra by mean and Gaussian curvatures. We can see on Fig-
formed by this barycenter, an edge of the mesh correspond- ure 8 that objects contains meaningful local informations
ing to the surface edge µ = µ1 associated to m, and the to characterise the geometrical variations. The mean curva-
barycenter of the points associated to these edges, and 3) ture is calculated on every points P of the mesh, according
the symmetric volume for the surface edge µ = µ2 (cf. Fig- to the usual formula, by summing the angles (θN )i of the
ure 7). The factor R1 corresponds to the P R level of resolu- normals of the dihedral formed by the adjacent faces to P
tion of ARG(M ), leading to the relation m∈M v(m) = 1 (i = 1, ..., Ne where Ne is the number of edges adjacent
as used for the relative area calculation a in [1]. Obviously, to P). Angles are weighted with the third of the mean area
this volume calculation method is valid only if triangles can (Sm )i of the two faces (barycentric area):
be oriented. Besides, an additive feature of the object sec-
3
tion associated to the node m could be the measure a(m)
 
1 X (θN )i
v(m)2 curvm (P) = 1 ,
which has no dimension. Ne i 3 (Sm )i
Figure 9. Colour histograms of the set of triangles
associated to an ARG node.

cally for each node of the ARG at the coarsest level of res-
olution, and then hierarchically and globally for each level
of resolution:
Figure 8. Mean curvature calculated on each ver- X
tex of the mesh (model from [25]). color∗ (m) = color∗ (n),
n∈{children of m}

where ∗ represents R, G, or B and color∗ (n) is the colour


The Gaussian curvature at a point P is calculated using the
histogram of the node n normalised as in Sections 4.3 and
angular defect formula:
7.4.
P
2π − j θj
curvg (P) = 1 P , 4.6. Objects similarity
3 j Sj
Assuming two objects M and N , the function used to
Sj being the area of the j th triangle having P as one of its calculate their similarity is obtained by summing the simi-
vertices, and θj is the angle at vertex P in this triangle. The larity function of the matching pairs of nodes {m, n}:
two histograms of these curvatures are calculated for each X
node at the finest resolution and afterwards propagated to SIM (M, N ) = 1 − sim(m, n),
the parents: {m,n}
X
curv∗ (m) = curv∗ (n), where SIM is minimal for similar objects. The 2D/3D fea-
n∈{children of m} tures presented in the previous sub-sections have been intro-
duced in the function sim(m, n), with weight parameters
where ∗ represents m or g, and curv∗ (n) is the histogram (λ1 , λ2 , λ3 , λ4 ) in accordance with the desired queries:
of the curvatures of the triangle surface associated to n and
1
normalised by the factor R1 · area(S) . As for the cords his- • λ1 · min(v(m), v(n)),
togram, each contribution to curv∗ is weighted by the as- • λ2 · min(cord(m), cord(n)),
sociated triangle area. It is also possible to use only one
histogram to represent the local curvature by combining the • λ3 · min(curv∗ (m), curv∗ (n)),
two curvatures (cf. Koenderink shape index [19, 8]) or to
use any other characterisation of local curvatures [20, 21]. • λ4 · min(color∗ (m), color∗ (n)).

4.5. Texture extraction We define the min calculation between two histograms
hist(m) and hist(n) as:
When browsing in a database of 3D textured objects, it
N bin
can be also useful to make queries upon the texture or the X
colour of the models. The joint use of shape and texture min(hist(m), hist(n)) = min(histi (m), histi (n)),
i=1
is a highly discriminative criterion, but is still not much
exploited in 3D. For this purpose, we use the information where Nbin is the size of the histogram and histi is the ith
included in the texture maps associated to the 3D objects. value.
These 2D features will naturally be merged to the 3D fea- For all objects M , N and O, the function SIM has the
tures previously presented. Our colour quantification is following properties:
done using three normalised histograms on (R, G, B) (cf.
Figure 9). The quantification of the texture is obtained lo- 1. Self-identity: SIM (M, M ) = 0,
2. Positivity: SIM (M, N ) ≥ 0, could match anywhere). We then propose a part match-
ing scheme which relies on the ideas presented in this arti-
3. Symmetry: SIM (M, N ) = SIM (N, M ), cle and which exploits the multiresolutional aspect and the
topological consistency criteria of the nodes for matching
As defined in [16], the function SIM is a semi-metric. The shape parts, and for extracting 2D/3D features to retrieve
triangle inequality is not verified as the function SIM takes the most similar node.
into account the pairs of matching nodes between two ob-
jects, and they can be different when comparing two dif- Part matching scheme
ferent objects to a same one. A new function SIM satis-
fying the triangle inequality would probably perform better Initialisation:
classification and searching [17] but remains to be found. 1. Choose an object M .
Finally, we have obtained a flexible multiresolutional 3D 2. Choose a level of resolution.
indexing scheme, both global and local, and using various 3. Choose the node m at this level of resolution
criteria illustrated by the different features extracted (topo- which corresponds to the part to be matched with
logical, geometrical, and colorimetric). other parts in the database.

Matching:
5. Application to part matching 4. Compute the similarity function between m and
every nodes n at the same level of resolution (or
Our shape matching method has both the properties of eventually at other close levels) among all the ob-
a high-level descriptor thanks to the global representation jects N in the database.
with the multiresolutional Reeb graph, and a low-level de- 5. Test topological consistency on nodes similar
scriptor with the local 2D and 3D features extracted at each to m with respect to a predefined threshold for the
level of resolution and for each node, especially at the finest similarity function.
level. This multiresolutional aspect can be exploited for 6. We consider now the subgraph of m and of all
many different purposes. Our current research work lead its children as the ARG of the part of interest, that
us to a promising application which consist on using the en- is the ARG of the truncated object region associ-
riched nodes for part matching as illustrated in Figure 10. ated to the node m. We consider similarly the sub-
As a matter of fact, the information included in the graph graphs of all the consistent nodes n as the ARGs
nodes are discriminant enough to be retrieved in a database. of the parts corresponding to nodes n. Then we
Moreover, the different representations of the nodes at each apply the general matching algorithm described in
level of resolution, which include their children nodes, can section 3 to these subgraphs in order to determine
ensure the precision of the matching. An approach of part the most similar parts to the one chosen in object
M as query.

6. Results
For each object, the ARG structure with the node at-
tributes are coded in a binary file. Its size is proportional
to the number of nodes and increases with the complexity
of the object. ARG files are computed once and require
∼10s for an object with 10000 triangles, the most time-
consuming being the computation of the function µ over
the object surface. For that we use the Dijkstra algorithm
with a complexity in O(N log(N )), N being the number
of vertices of the mesh. The ARG files stand for the fea-
Figure 10. Nodes taken from Reeb graphs for part ture vectors in the indexing framework, and the comparison
matching (models from [23]). of two objects takes between 1 to 2 millisecond with a PC
P4-2GHz. The 3D model database was created by recon-
structing museum objects [18, 22], and downloading objects
matching was presented in [14], but the skeletal represen- from the Internet [23, 24, 26]. Most of the models have a
tation of the objects was not well adapted for this purpose. high resolution (10000 to 50000 triangles) and closed sur-
The lack of a multiresolutional strategy is obvious as a part face with oriented triangles. To show the advantages of the
of a skeleton could be a too coarse description (a cylinder Augmented Reeb Graph, we have made our similarity tests
between eight different classes of objects (cf. Figure 11). other objects. For our classification test (cf. Figure 12), we
Several of these classes could be easily confused by com- used the values αl = 0.3, βl = 0.15, γl = 0.15, δl = 0.2
and l = 0.2 with L1 norms for the loss function, and
αs = 0.65 and βs = 0.35 for the sim function. A test
using the original approach of [1] was performed to illus-
trate the effectiveness of our approach (cf. Figure 13). We

Figure 12. Results of the similarity function on 50


models using the ARG matching method: similar
models are retrieved and classes are well sepa-
rated.

Figure 11. Eight classes of objects were used for


our tests. There are different numbers of objects Figure 13. Results of the similarity function us-
per class for a total of 50 objects (models from [22, ing the original approach in [1]: more unwanted
23, 24, 26]). similarity are found between objects of different
classes.

mon approaches: statues, human models, elongated vases,


etc. Some statues are representations of human beings but use the weight αl = αs = 0.5 and βl = βs = 0.5. As said
don’t look like the models of the human class, some others previously, the lack of geometrical information compared
are very unique in the database and should not match with to our extension leads to less nodes matching and then sim-
ilar objects could be missed when the topology matching is
wrong.
These preliminary results reveal that the ARG method is
more efficient to retrieve object classes. The six other tested
methods make a strong confusion between statues and other
objects (cf. Figure 14). Those with the best results (com-
bined histograms of cord lengths and cord angles with the
1st principal axis [5] and 3D Hough Descriptor [8]) are
more noisy than the ARG and still present some class mix-
ing and wrong matching. Moreover, the results concerning Figure 15. Reeb graphs of different objects at
the classification of the vases show that they are not efficient level of resolution 4: the graphs keep the genus
enough to distinguish the difference between vases with one (models from [22, 23]).
or two handles, whereas the ARG is. The D2 shape distri-
bution [6], the complex EGI [10], the 3D shape spectrum
descriptor [8] and the surface/volume ratio [4] were not ef-
ficient on our database. The Figure 15 illustrates the Reeb 7. Conclusion
graph for objects of different classes and obviously its ha-
bility to catch the genus of the object and then its specific Keeping in sight our problematic of textured 3D object
power for graph matching. indexing and content-based retrieval, our study leads us to
enhance the approach presented in [1] by introducing the
Augmented Reeb Graph.
Geometrical and colorimetric attributes are added to the
topological consistency criteria which are also extended.
The loss function calculation is consequently modified, and
the similarity function is improved by merging the different
2D/3D features (geometrical and visual).
Then, we obtain a multiresolutional and multicriteria shape
descriptor. Our results demonstrate the ability of the
ARG to retrieve similar objects classes in a database, and
particularly objects with long appendices, or with the same
genus. However this method seems not so well adapted
for compact objects as busts or faces. And actually the
function µ works with meshes that have only one connected
component, thus the objects should have been correctly
modelized.
Further works could be done to improve the characterisa-
tion of the texture. For example by unfolding the texture
associated to each node (patch), and reparametrizing it
to extract spatial or frequency informations. Our method
aims the interactivity and the flexibility, but requires lots
of parameters (weight terms of the different features). It
could be then interesting to apply optimisation techniques
on parameters to define default values in case of specific
applications for which learning sets could be provided.
Perspectives of research also concern the choice of the
function µ and the construction of the Reeb graph. It
seems that associating methods using a robust graph (which
allows topological perturbation) such as the Extended Reeb
Graph [15] with our approach of Augmented Reeb Graph
(with its multiresolutional aspect), could be very efficient.
The similarity function can be a useful tool to compare
Figure 14. Results of the six other tested meth- objects with similar shapes, especially when associated to
ods: there are lots of noise and classes are mixed. 2D and 3D features. For example, they could establish
measurements between two objects depending on their vol-
ume, curvatures or visual differences. Moreover, it is also [9] B.K.P. Horn. Extended Gaussian Image, Proc. of the IEEE,
possible to use these informations to make classifications Vol.72, pp.1671-1686, 1984.
with respect to any of the criteria offer by the extracted
[10] S. B. Kang, K. Ikeuchi. The complex EGI: a new representa-
features (colour,size,...). In fact, as the technology of the tion for 3D pose determination, IEEE Trans on PAMI, Vol.16,
computer keeps on growing (CPU, memory and HD size), No.3, pp.249-258, Mar. 1994.
Augmented Reeb Graph will allow us to integrate and
merge all of the future most relevant shape descriptors. [11] J.A. Goldak, X. Yu, A. Knight, L. Dong. Constructing dis-
Also, other applications such as part matching seems crete medial axis of 3-d objects, International Journal of
Computational Geometry and Applications, 1(3), pp.327-339,
accessible with our approach but has not been developed
2001.
yet and remains to be fully tested. Finally, we would like
to point out again the flexible and robust aspects of our [12] F. Leymarie and B. Kimia. The Shock Scaffold for Rep-
approach, its effectiveness and accuracy relying on the resenting 3D Shape, Visual Form, Springer-Verlag, Lecture
fusion of different techniques. Notes in Computer Science, no.2059, pp.216-229, 2001.

[13] A. Verroust, F. Lazarus. Extracting Skeletal Curves from 3D


Scattered Data, The Visual Computer, Springer, 16(1), pp.15-
Acknowledgements 25, 2000.

This work has been partially supported by the SCULP- [14] H. Sundar, D. Silver, N. Gagvani, S. Dickinson. Skeleton
TEUR European project IST-2001-35372 [22]. The authors Based Shape Matching and Retrieval, Shape Modeling Inter-
would like to thank Pierre-Alexandre Pont for having coded national, Seoul, Korea, pp.130-142, May 2003.
the algorithm of [1] and for the additional topological con- [15] S. Biasotti, S. Marini, M. Mortara, G. Patanè, M. Spagn-
sistency criterion, and Bianca Falcidieno and Marco Attene uolo, B. Falcidieno. 3D Shape Matching through Topologi-
for their precious advice concerning the function µ. cal Structures, 11th Discrete Geometry for Computer Imagery
conference, Springer-Verlag, Lecture Notes in Computer Sci-
References ence, no.2886, 2003.

[16] J.W.H. Tangelder and R.C. Veltkamp. Polyhedral Model


[1] M. Hilaga, Y. Shinagawa, T. Kohmura, T.L. Kunii. Topology
Comparison Using Weighted Point Sets, Shape Modeling In-
Matching for Fully Automatic Similarity Estimation of 3D
ternational, Seoul, Korea, pp.119-129, May 2003.
Shapes, ACM SIGGRAPH, Los Angeles, CA, USA, pp.203-
212, Aug. 2001. [17] J. Barros, J. French, W. Martin, P.Kelly, and M. Cannon. Us-
ing the triangle inequality to reduce the number of compar-
[2] Y. Shinagawa, T.L. Kunii and Y.L. Kergosien. Surface Cod-
isons required for similarity-based retrieval. Proc. of SPIE,
ing Based on Morse Theory, IEEE Computer Graphics and
vol.2670, pp.392-403, 1996.
Applications, Vol.11, No.5, pp.66-78, Sept. 1991.
[18] C. Hernández Esteban and F. Schmitt. Silhouette and Stereo
[3] D.V. Vranic, D. Saupe and J.Richter. Tools for 3D-object re-
Fusion for 3D Object Modeling, 3DIM, 4th Int. Conf.
trieval: Karhunen-Loeve Transform and spherical harmon-
on 3D Digital Imaging and Modeling, pp.46-53, 2003.
ics, IEEE Workshop Multimedia Signal Processing, Cannes,
http://www.tsi.enst.fr/3dmodels
France, pp.293-298, Oct. 2001.
[19] J. Koenderink. Solid Shape, The MIT Press, 1990.
[4] C. Zhang and T. Chen. Efficient Feature Extraction for 2D/3D
Objects in Mesh Representation, ICIP, Thessaloniki, Greece, [20] D. Cohen-Steiner, J-M. Morvan. Approximation of the Cur-
pp.935-938, Oct. 2001. vature Measures of a Smooth Surface endowed with a Mesh,
Rapport de recherche INRIA, Thème 2, no.4867, Jul. 2003.
[5] E. Paquet and M. Rioux. Nefertiti: a Query by Content Soft-
ware for Three-Dimensional Models Databases Management, [21] Takeshi Masuda. Surface Curvature Estimation from the
3-D Digital Imaging and Modeling, Proc. Int. Conf. on Recent Signed Distance Field, 3DIM, 4th Int. Conf. on 3D Digital
Advances, pp.345-352, 1997. Imaging and Modeling, pp.361-368, 2003.
[6] R. Osada, T. Funkhouser, B. Chazelle, and D. Dobkin. Shape [22] http://www.sculpteurweb.org
Distributions, ACM Trans. on Graphics, 21(4), pp.807-832,
Oct. 2002. [23] http://www.fantasticarts.com

[7] K. Kazhdan and T. Funkhouser. Harmonic 3D Shape Match- [24] http://www.3dcafe.com


ing, SIGGRAPH Technical Sketches, p.191, July 2002.
[25] http://graphics.stanford.edu/data/3Dscanrep/
[8] T. Zaharia, F. Prêteux. Indexation de maillages 3D par de-
[26] http://shape.cs.princeton.edu/benchmark/
scripteurs de forme, RFIA, pp.48-57, Angers, France, Jan.
2002.

You might also like