IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICSPART C: APPLICATIONS AND REVIEWS, VOL. 37, NO.
6, NOVEMBER 2007 1081
Content-Based 3-D Model Retrieval: A Survey
Yubin Yang, Member, IEEE, Hui Lin, Member, IEEE, and Yao Zhang
AbstractAs the number of available 3-D models grows, there
is an increasing need to index and retrieve them according to their
contents. This paper provides a survey of the up-to-date methods
for content-based 3-D model retrieval. First, the new challenges
encountered in 3-D model retrieval are discussed. Then, the sys-
tem framework and some key techniques of content-based 3-D
model retrieval are identified and explained, including canonical
coordinate normalization and preprocessing, feature extraction,
similarity match, query representation and user interface, and
performance evaluation. In particular, similarity measures using
semantic clues and machine learning methods, as well as retrieval
approaches using nonshape features, are given adequate recog-
nition as improvements and complements for traditional shape-
matching techniques. Typical 3-D model retrieval systems and
search engines are also listed and compared. Finally, future re-
search directions are indicated, and an extensive bibliography is
provided.
Index Terms3-D object, 3-D shape, content-based retrieval,
feature extraction, query interface, similarity match.
I. INTRODUCTION
S AN EXPANSION from audio (1st wave of multime-
A dia), image (2nd wave of multimedia), and video (3rd
wave of multimedia), 3-D digital models and 3-D scenes are
now regarded as the 4th wave of multimedia and are suffi-
ciently realistic and complex to convey an increasing amount of
perceptual detail [1]. Unlike 2-D images, 3-D models are able
to overcome the illusion problem caused by the human eye;
thus object segmentation becomes less error prone and easier
to achieve [2]. Modern computer technology, powerful comput-
ing capacity, combined with new acquisition and new modeling
tools make it much easier and cheaper to create and operate 3-D
models with basic hardware, which consequently produces an
increasing amount of 3-D models in various sources, such as pro-
fessional 3-D model collections (e.g., Biology, Medicine, Chem-
istry, Archaeology, and Geography) and on the Internet [3]. This
naturally raises an increasing need for new retrieval tools, by
which these large-scale and complicated new generation me-
dia can be easily organized and found by computer users. In
addition, modeling highly realistic 3-D models is still a very la-
borious, high-cost, and time-intensive process. If the 3-D models
Manuscript received February 2, 2005; revised February 20, 2006 and May
9, 2006. This work was supported in part by the Key Program of the National
Natural Science Foundation of P.R. China under Grant 60723003, in part by
the National Natural Science Foundation of P.R. China under Grant 60505008,
and in part by the Natural Science Foundation of Jiangsu Province under Grant
BK2007520. This paper was recommended by Associate Editor L. Zhang.
Y. Yang is with the State Key Laboratory for Novel Software Technology,
Nanjing University, Nanjing 210093, China. and also with the Joint Laboratory
for GeoInformation Science, Chinese University of Hong Kong, Shatin, N.T.,
Hong Kong (e-mail: yangyubin@ieee.org).
H. Lin is with the Joint Laboratory for GeoInformation Science, Chinese
University of Hong Kong, Shatin, N.T., Hong Kong.
Y. Zhang is with the Department of Information Management, Nanjing
University, Nanjing 210093, China.
Digital Object Identifier 10.1109/TSMCC.2007.905756
1094-6977/$25.00 2007 IEEE
currently available can be efficiently retrieved and reused, much
less time and effort will be needed to complete the modeling
task [4].
Content-based is a preferred method for handling multime-
dia data efficiently [5][7]. However, compared with the thriv-
ing development and achievements in 1-D and 2-D multimedia
search, the 3-D model search is still in its infancy. Many Web-
sites only allow users to search 3-D models in quite a limited and
primitive way, such as browsing a directory structure, running
a text keyword search [8][11], or retrieving by file type and
file size [12]. Thoese traditional text-based search techniques
are not effective for 3-D models, as they suffer from problems,
such as low efficiency, low accuracy, and high ambiguity [6],
[7]. Moreover, the most significant problem is that 3-D models
contain both shape and appearance information, which is hard
to represent and query, using text keywords alone.
To address these issues, the idea of retrieving 3-D models in a
Content-based scheme has already attracted considerable at-
tention as a new hotspot in several research communities, such
as computer vision, computer graphics, geometric modeling,
pattern recognition, mechanical computer-aided design (CAD),
and molecular biology [13]. This Content-based scheme is
now developing into a Content-based 3-D Model Retrieval
methodology, concentrating on the representation, recognition,
and matching of 3-D models on the basis of extraction and
comparison of intrinsic representative features, such as shapes,
colors, textures, and light distributions [14][21]. A growing
number of researchers have become involved in this area and
have already made much progress [22][25]. Many algorithms
have been proposed and reported, and there has been subsequent
increase in the publication of academic papers and books related
to this topic, in a wide range of international journals and con-
ferences [13], [14], [16], [17], [26][30]. The new international
standard MPEG-7 has also covered some 3-D object descriptors
as one of its feature sets [5], [31], [32]. In fact, content-based
3-D model retrieval can be applied widely in many domains,
such as CAD, cultural heritage applications, robotics, molecu-
lar biology, virtual geography environment (VGE), 3-D spatial
terrain, medicine, chemistry, military, and industrial manufac-
turing [26], [33][40], [55]. It can also be potentially applied in
e-Business and Web search engines in distributed data environ-
ments [14], [41][44].
In this paper, we review the state-of-the-art research methods
on content-based 3-D model retrieval and explore the important
issues and the outstanding challenges of this area. Following
this introduction, the remainder of the paper is organized as
follows. In Section II, the basic concepts of content-based 3-D
model retrieval are briefly introduced, highlighting the new chal-
lenges it faces. This is followed by a description of the archi-
tecture framework, leading in Section III to a discussion of
the canonical coordinate normalization process and 3-D model
1082 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICSPART C: APPLICATIONS AND REVIEWS, VOL. 37, NO. 6, NOVEMBER 2007
preprocessing technologies. Section IV provides an appropriate
categorization of the currently available feature representa-
tion and extraction technologies. Section V reviews the rep-
resentative work in similarity match techniques. Section VI
briefly outlines the query representation and user interface tech-
nologies. Section VII discusses performance evaluation meth-
ods. Section VIII enumerates some typical content-based 3-D
model retrieval systems and search engines. Finally, Section IX
concludes the paper with a number of directions for future
research.
II. CONTENT-BASED 3-D MODEL RETRIEVAL
A. Overview
The fundamental processing flow of content-based 3-D model
retrieval can be described as follows: first, the compact and de-
scriptive features, such as geometric shapes, spatial and topo-
logical relationships, statistical measures, textures, and material
colors, are computed and extracted automatically from the 3-D
data to build the multidimensional information indices. Then,
the similarity measure between a query and each target model
in the database is defined in the multidimensional feature space.
The similarity degrees are then calculated and sorted so that
the models having the nearest similarity are returned as the
matched results, on the basis of which, browsing and retrieval
in 3-D model databases are finally implemented [14]. Here,
Content-based means that 3-D models are retrieved accord-
ing to their visual feature indices, which are automatically (or,
semiautomatically) extracted and expected to characterize their
contents.
The ultimate goal of content-based 3-D model retrieval is to
approach human visual perception so that the semantically simi-
lar 3-D models can be correctly sourced by their looks. However,
the types of methods outlined earlier, which can be termed, low-
level similarity-induced semantics, capture some, but not all as-
pects of the content of a 3-D model, and do not coincide with the
high-level information it contains. For example, a sphere-like
shape can be used to describe a 3-D ball, but how can we judge
whether it is a 3-D ball or a 3-D model of the globe by using
that feature alone? This is the well-known semantic gap prob-
lem [45] which indicates the relatively limited descriptive power
of low-level visual features for approximating subjective high-
level perception. Therefore, extraction of high-level features
which can derive semantics from low-level features should also
be integrated as an important part in content-based 3-D model
retrieval.
B. New Challenges
Content description and retrieval tasks have become much
more complex and difficult due to the new intricacy existing in
the features of 3-D models, leading to new challenges, listed as
follows.
First, building correct feature correspondence for 3-D models
is more difficult and time-consuming. The 3-D models possess
more complex and excessive poses than 2-D media, with differ-
ent translations, rotations, scales, and reflections. This gives 3-D
models many more arbitrary and unpredictable positions, orien-
tations, and measurements and makes the 3-D models difficult to
parameterize and search [2]. However, it is necessary to search
3-D models invariantly with respect to translation, rotation, scal-
ing, and reflection [46]. Therefore, in many cases, more ad-
ditional alignment-normalization (pose registration) processes
may be required to align 3-D objects to their canonical coordi-
nate frame, or more intricate mappings or transformations for
extracting invariant feature representations of a 3-D model be-
fore a similarity match, which are time-consuming, computing-
intensive, and unstable.
Second, the diversity of the representations of 3-D infor-
mation may impede the implementation of simple convenient,
and efficient 3-D model retrieval systems. There is no single
common 3-D shape format that serves as a standard. The 3-D
models are usually represented with two kinds of data: geo-
metric properties and appearance attributes. Geometric proper-
ties have a wide variety of representations including different
kinds of information like point data, surface data, volumet-
ric data, and structures, such as solid structures used in the
CAD/computer-aided manufacturing (CAM) field, parametric
surfaces orpolygon meshes, implicit surfaces, volumetric ar-
rays of voxel grids, or just unstructured polygon-soup and
point clouds, while appearance attributes may contain mate-
rial types, material colors, transparency, reflection coefficients,
and texture mapping. Due to the representation multiplicity,
most of the currently available 3-D model matching algo-
rithms merely depend on 3-D shape properties based on some
specific 3-D data formats. How to overcome the unnecessary
complexity and noneffective matching caused by format di-
versity is one of the major challenges of content-based 3-D
model retrieval. To identify viable solutions to address these
issues, it is necessary to develop some new types of high-
level descriptors to provide a unified view on perceptual un-
derstanding of a 3-D model. Nevertheless, a 3-D model usually
lacks high-level semantic clues. Therefore, there also are chal-
lenges for establishing effective mapping mechanisms between
low-level 3-D data representations and high-level semantic
descriptions.
Third, 3-D data representations have been designed for ef-
ficient visualization tasks, resulting in many problems for fea-
ture indexing and similarity comparison. For example, some
3-D representations are not inherently well-defined, such as vir-
tual reality modeling language (VRML)-like polygon soup
and some unclosed meshes. It is difficult and ineffective to
transform these into well-defined representations before feature
extraction [47][49]. Therefore, to accept polygon soup and
other ill-defined 3-D models is a further challenge for 3-D model
retrieval tasks [50].
Finally, 3-D models have both considerable appearance at-
tributes and complex geometric properties, which greatly in-
creases the amount of content information. In addition, the di-
mension of 3-D data is also too high to be dealt with effectively
and efficiently. Moreover, the multiresolution feature represen-
tations should be effectively generated in order to be robust
against different level-of-detail (LOD) of 3-D model represen-
tations [46], [51].
YANG et al.: CONTENT-BASED 3-D MODEL RETRIEVAL: A SURVEY
Fig. 1. Typical architecture framework of content-based 3-D model retrieval [3], [52], [53].
C. Architecture Framework
A typical content-based 3-D model retrieval system gener-
ally consists of four main components: 1) canonical coordi-
nate normalization (pose registration) and model preprocessing;
2) feature extraction: the generation of both low-level features
capturing the 3-D shape, appearance properties, and high-level
semantic features; 3) similarity match: relevance ranking pro-
cedure according to similarity degrees calculated; and 4) query
interface: a practical online user interface designed to represent
and process user queries. A typical architecture framework is
shown in Fig. 1.
Different from the traditional 3-D object recognition tasks,
which are usually performed at the cost of high computational
complexity by establishing correspondences between a pair of
3-D models and then comparing them, content-based 3-D model
retrieval needs to be performed on a per-model basis, which
means the feature used for matching should be calculated and
stored independently of the target 3-D models [54]. This also
allows for the so-called offline feature extraction process be-
cause there is no need to explicitly establish correspondence.
Therefore, for the genuine online retrieval phase, matching is
performed by comparing the values of feature descriptors for
each 3-D model. As can be seen from Fig. 1, the feature in-
formation of each 3-D model is extracted from the 3-D model
database during an offline stage to enable comparison to online
queries later on.
III. CANONICAL COORDINATE NORMALIZATION
AND PREPROCESSING
A. Overview
In many cases, we first need to normalize the size and orien-
tation of a 3-D model before extracting its features. The normal-
ization step aims to ensure that the same feature representation
can be properly extracted from a 3-D model with any differ-
ent scale, position, and orientation. This allows us to perform
search and retrieval tasks on a per-model basis, without fur-
ther alignment of 3-D models to each other. At present, there
are two approaches to realize such a per-model based normal-
ization [20]: 1) Normalization: finding a canonical coordinate
frame using methods similar to those used in principle com-
ponent analysis (PCA), also referred to as pose estimation
1083
or pose registration, 2) Invariance: defining and extracting
feature descriptors that possess the inherent invariance charac-
teristics, so as not to change under any rigid transformations. The
Invariance approaches have been accorded increasing weight
in recent research because of their robustness and simplicity.
However, the invariance they possess is not always complete
and all-sided. Moreover, the computation of these feature de-
scriptors is necessarily completed over a unit coordinate frame.
Hence, to guarantee the descriptive power and robustness of
the feature representations, the canonical coordinate normaliza-
tion, such as alignment and scaling, is also necessary before
invariance-keeping features are extracted.
Besides the normalization process, performing some pre-
processing steps on 3-D models before their features are ex-
tracted is inevitable. These steps include the transformation
between different 3-D data representations (e.g., to transform
polygon meshes into voxel grids), the partition of model units,
and vertex clustering, etc. [3], [14], [55].
B. Normalization Steps
1) Translation: First, the models center of mass should
be shifted to the coordinate origin, i.e., I1 = I c = {u|u =
v c, v I}, where I is the original coordinate, I1 is the new
coordinate after translation, and c is the models centroid [20].
2) PCA-Based Rotation: Next, for the rotation invariance,
the most popular approach is the PCA [20], [56], [57]. PCA
is used to determine the canonical coordinate system axes of a
3-D model, based on calculating the corresponding eigenvectors
and the resulting diagonal matrix R of eigenvalues, decreasingly
ordered by their values. The rotation transformation is repre-
sented as: I2 = R I1 = {x|x = R u, u I1 }, where I1 is
the 3-D models coordinate frame before rotation and I2 is the
new coordinate frame after rotation, which are identical to the
directions having the top three largest variances of the point
distribution.
The general PCA transformation in 3-D model retrieval is
defined on the given set of representative points of a 3-D model,
such as vertices, centroids of each surface, or even randomly
selected locations on each surface using statistical techniques
(e.g., the Monte Carlo approach) [58]. In considering the differ-
ent sizes of triangles or meshes of a 3-D model, some appropri-
ate weighting factors, proportional to their surface area can be
1084 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICSPART C: APPLICATIONS AND REVIEWS, VOL. 37, NO. 6, NOVEMBER 2007
accommodated, so as to make the transformation more robust
and improve the reliability and veracity of feature representa-
tion [20], [59], [60]. However, the point-based PCA transfor-
mation may cause an inaccurate normalization result that will
seriously affect the retrieval precision if the chosen vertices do
not distribute evenly on the surface. Therefore, a more thor-
ough improvement, termed CPCA (continuous PCA), which
performs PCA transformation based on the whole 3-D poly-
gon mesh, is proposed in literature [57]. CPCA generalizes the
PCA transformation by using the sums of integrals over sur-
faces instead of the sums over selective vertices. Assume that
the whole size
of all the surfaces
 in a 3-D model is repre-
sented as S = M i=1 S i = I dv, where v I, I = m i=1 Ti is
the point set of the 3-D model, and T = m i=1 T i (T i R3 ) is
the triangleset. The covariance matrix R is then defined as
R = 1/S I v v T dv.
The PCA algorithm is fairly simple and efficient. However, it
may erroneously assign the principal axes and produce inaccu-
rate normalization results, especially when the eigenvalues are
equal or close to each other, which usually happens to different
models within the same category [14], [22].
3) Reflection and Scaling: A diagonal flipping matrix F is
designed to accomplish the reflection invariance, which ensures
a model and its reflection will have the same feature descriptor.
Finally, the 3-D model should also be scaled by multiplying a
proper scaling coefficient s to a certain unit size to guarantee the
scaling invariance. The definition of the flipping matrix and scal-
ing coefficient can be found in the literature [57]. Consequently,
the whole normalization process is illustrated as follows [20]:
(I) = s1 F R (I c). (1) rection vectors and set rays along each direction vector from
IV. FEATURE EXTRACTION
The new adopted features in content-based 3-D model re-
trieval include 2-D shape projections, 3-D shapes, 3-D appear-
ances, and even high-level semantics, which are required not
only to be extracted, represented, and indexed easily and ef-
ficiently, but also to effectively distinguish the similar models
from the dissimilar models, invariant to typical affine transfor-
mations [20], [61][63].
According to the different aspects of the content they repre-
sent, the features of a 3-D model can be roughly categorized
into two main types: 1) shape features, namely, geometry and
topology features and 2) appearance features, which represent
some important cognitive characteristics such as material col-
ors, reflection coefficients, and textures mapping. Categories and
comparisons of these feature extraction methods are provided
in Table I in detail.
A. Shape Representations and Extraction
Currently, most of the work on shape feature extraction places
emphasis on geometrical and surface topological properties of
3-D shape features, based on surfaces, voxels, vertex sets, and
structural shape models [2], [13], [20], [29], [30], [42], [52].
Generally, geometrical features usually represent the specific
shape and spatial position of surfaces, edges, and vertices, while
topological features maintain the linking relationship between
surfaces, edges, and vertices. A review of the existing catego-
rizations of 3-D shape representations according to different
criteria and classification levels is listed in Table II [13], [29],
[30], [52].
In this paper, in order to provide a more unified perspective
on the way in which shape characteristics of a 3-D model are
represented, a new general taxonomy of 3-D shape feature ex-
traction methods is proposed the salient features of which are
as follows: 1) approaches based on global geometrical analy-
sis; 2) approaches based on function mapping; 3) approaches
based on statistical properties computation and; 4) approaches
based on topology analysis. However, it should be borne in mind
that these methods are not absolutely independent and isolated.
In fact, many of them are quite interdependent. The purpose
of our taxonomy is to provide a rational and comprehensi-
ble classification and summarization on the existing research
literature.
1) Global Geometrical Analysis Methods: The global ge-
ometry of a 3-D model is analyzed by directly sampling the
vertex set, the polygon mesh set, or the voxel set in the spatial
domain. Aspect ratio, binary 3-D voxel bitmap, and 3-D angles
of vertices or edges may be considered the most simple and
straightforward [64], although their discriminative powers are
limited. These types of analyses generally use PCA-like meth-
ods to align the model into a canonical coordinate frame at first,
and, then, define the shape representation on this normalized ori-
entation. For instance, Vranic et al. [20] proposed a ray-based
geometrical feature representation. They sampled a 3-D model
in its canonical coordinate frame as a set of regular spaced di-
the coordinate origin, which intersected with the triangle mesh
of a polyhedron surrounding the 3-D model. For each direction,
the maximum distance from the intersected triangle mesh to the
coordinate origin was computed and all the distance samples
composed a feature vector. Using a similar idea, Yu et al. [65]
extracted the 3-D global geometry as distance map and surface
penetration map features.
Suzuki et al. [66] proposed a geometrical feature representa-
tion based on the concept of equivalence class (see Fig. 2).
First, they divided a 3-D model into a set of grid cells and com-
puted the number of vertices (or normal vectors of a surface) in
each grid cell, which was divided into different equivalence class
groups. The equivalence class was defined as a set of grid cells
sharing the common characteristics, for example, the vertices,
whose new position was still in the same set. The feature vector
was then calculated according to the properties of equivalence
classes, such as their vertex numbers. Heczko et al. [60] and
Vranic et al. [68] both implemented an octree structure method
to represent the shape features of 3-D volumetric models by
fulfilling a multiresolution subdivision of the 3-D model space.
For each grid cell, they took the sum of mesh sizes bounded
by the grid cell, or the 3-D Fourier transformation coefficients
as the feature components, which formed a feature descrip-
tor of 2r 2r 2r dimensions, where r was the resolution of
octree representation. Tangelder et al. created a method using
weighted point sets as the shape descriptor for a 3-D polygon
YANG et al.: CONTENT-BASED 3-D MODEL RETRIEVAL: A SURVEY
TABLE I
CATEGORIES AND COMPARISONS OF 3-D FEATURE EXTRACTION METHODS
mesh [69]. They also enveloped the object in a 3-D voxel grid
and represented the shape as a weighted point set by selecting
one representative point for each nonempty grid cell. They then
selected the vertex with the highest Gaussian curvature or the
area-weighted mean of all the vertices in a grid cell, to represent
the models geometry features. The experimental results given
by Tangelder were very promising, but their main shortcoming
was the long time it took to compute the descriptors. As for
3-D industrial solid models, Cicirello et al. [70] and McWherter
et al. [18] both compared 3-D shapes by extracting the geomet-
rical and engineering features of 3-D models in spatial domains.
The common characteristic of these methods is that they are
almost all derived directly from the elementary unit of a 3-D
model, that is, the vertex, polygon, or voxel, and a 3-D model
is viewed and handled as a vertex set, a polygon mesh set, or a
voxel set. Their advantages lie in their easy and direct deriva-
tion from 3-D data structures, together with their relatively good
representation power. However, the computation processes are
usually too time-consuming and sensitive for small features.
Also, the storage requirements are too high due to the dif-
ficulties in building a concise and efficient indexing mecha-
nism for them in large model databases [71], [72]. In order
to improve the overall performance, the Divide-and-Conquer
strategy was adopted in the feature extraction process. In some
1085
cases, the low efficiency was mainly caused because some of
the feature representations cannot be computed directly from
the 3-D meshes, which need to, first, be transformed into a 3-D
voxel space. This process is time-consuming and requires a large
amount of storage space [73][75]. To address this issue, Zhang
et al. [76] proposed a global geometrical analysis algorithm us-
ing Divide-and-Conquer strategy, with no need of volumetric
transformation. They first computed the features for each el-
ementary surface (a triangle or a tetrahedron) of a 3-D mesh
model, and then summed them up to form the global feature
vector.
2) Function Mapping Methods: Function mapping methods
establish a functional mapping from the original 3-D model to
a predefined domain, typically a spherical function with invari-
ance properties or several representative 2-D planar views with
reduced dimensions. This has long been studied in 3-D engi-
neering design and CAD communities, and has become one of
the popular means to extract 3-D shape signatures.
a) Spherical Mapping: The first viable method is to map
arbitrary 3-D shapes into a series of spherical functions to guar-
antee rotation invariance. Horn et al. [77] generated extended
Gaussian image (EGI) mapping from a 3-D object to a Gaus-
sian sphere, which associated a point on the 3-D surface with the
corresponding point on the Gaussian sphere, by finding point
1086 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICSPART C: APPLICATIONS AND REVIEWS, VOL. 37, NO. 6, NOVEMBER 2007
TABLE II
REVIEW OF DIFFERENT CATEGORIZATIONS OF 3-D SHAPE REPRESENTATIONS
Fig. 2. Feature extraction based on equivalence class [67].
Fig. 3. Multiresolution 3-D model reconstruction using spherical harmonic
mapping [79].
pairs having the same surface normal vector. The distribution of
those surface normal vectors on the Gaussian sphere was then
extracted to match 3-D surfaces. Similar to the EGI feature,
Delingette et al. [78] presented a representation termed, Sim-
plex Angle Images (SAI). A 3-D object was first approximated
as a discrete ellipsoid using a surface deformation algorithm,
then mapped to a normalized sphere. The curvatures of every
surface finally served as shape signatures.
Saupe et al. [79] adopted a spherical harmonic function as the
basic mapping function in the feature extraction process. A unit
sphere S 2 centring at the centroid of a 3-D object was defined
then as a function mapping r(u)(u S 2 ) (as shown in Fig. 3),
which was given as follows:
r : S 2 C, r(u) = x(u) + i y(u).
They then performed the discrete Fourier transform (DFT) on
a number of samples of x(u) along some radial directions uij
in the spherical coordinate and took the first K transforming
coefficients as the feature vector. To improve on this method,
Vranic et al. [46] calculated the spherical Fourier transforming
coefficients by directly using the spherical harmonics function
in frequency domain instead of sampling the 3-D surface first.
Funkhouser et al. [14], [62] also introduced a rotation-invariant
spherical harmonic representation, whose key idea is to establish
a spherical function to depict the energies at different frequen-
cies. This approach can be viewed as an expansion of the Fourier
descriptor shape representation [80] to spherical functions. A
3-D volumetric model space was first represented as a series of
spherical harmonics by mapping and decomposing into a set of
concentric spheres with different radii. Then, those harmonics
within each frequency were summed up to compute the norm
of each frequency component, i.e., the spectrum of energies of
the spherical function as 3-D shape signatures. Fig. 4 illustrates
an instance of this calculation process. Those signatures are in-
variant and dimension-reduced so as to save both storage space
and matching time. Also, using the concept of energy, Leifman
et al. [81] proposed a shape feature representation based on
a sphere projection, which computed the amount of energy
required to deform a 3-D model into a predefined 3-D sphere
whose origin was located in the 3-D models centroid.
From a different perspective, Kazhdan et al. [61], [82] intro-
duced a reflective symmetry descriptor that captures the sym-
metry measures of 3-D objects. For each 2-D plane crossing
through the 3-D models centroid, a symmetry distance was de-
fined to compute how similar the model was to its reflection.
The 3-D voxel grid was then decomposed into a collection of
concentric spheres so that all the symmetry distances on those
spheres were computed. Finally, all the different values of those
symmetry distances were combined to form a reflective sym-
metry descriptor as a 3-D shape feature.
The methods described earlier produce invariant shape fea-
tures, which avoids the time-consuming canonical coordinate
normalization process in feature extraction. However, they also
have some shortcomings. Firstly, it is generally assumed that
a 3-D model will have valid topology (for meshes), or explicit
volume (for volumetric models), which cannot be guaranteed
in practice. Secondly, the spherical function mapping process is
complicated and time-consuming.
b) 2-D Planar Mapping: The other type of function map-
ping is quite intuitive and straightforward. Several 2-D func-
tional projections of a 3-D model or 2-D planar views from
different perspectives are generated and combined as shape or
silhouette feature descriptors.
The 2-D functional projection reduces the 3-D matching prob-
lem into a 2-D case without computing multiple views of the
object. Johnson et al. [26] proposed a spin image representa-
tion, a 2-D descriptive image associated with a sampling vertex
set on a 3-D surface, for which both the position and direction
information were involved. The x and y coordinate values of
the 2-D spin image were defined as the accumulated values of
(2) two different distance functions of the 3-D vertices, and the
correlation coefficient between two spin images was computed
as the similarity measure. However, since a 3-D model usually
consists of many surfaces, there is a large set of spin images
generated for each 3-D model. To achieve more concise and
compact feature representation, the original set of spin images
is compressed using the PCA method. Zhang et al. [83] reduced
the 3-D surface matching problem to a 2-D image matching
problem by employing harmonic map theory [84], which stud-
ies the boundary mappings between different metric manifolds
in terms of the energy-minimization principle. Gu et al. and
YANG et al.: CONTENT-BASED 3-D MODEL RETRIEVAL: A SURVEY 1087

Fig. 4. Sphere harmonic representation [14].

Praun et al. [85], [86] discussed the geometry image concept,

a simple 2-D array of quantized points with useful attributes,
such as vertex positions, surface normals, and textures. Laga
et al. [87] applied this method to 3-D shape matching by simpli-
fying the 3-D matching problem to measure similarity between
parameterized, 2-D geometry images. All those methods make
use of specific 3-D geometry information from a 3-D model in
their 2-D mapping process.
Compared with these methods, the 2-D mapping methods
that establish mappings from a 3-D view to a set of specific
2-D planar views from different angles are much more natural
and simple. The basic idea is that if two 3-D shapes are simi-
lar, they should be similar from many different views. Thus, 2-D
shapes, such as 2-D silhouettes, can be extracted and adopted for
3-D shape matching. There is a prolific amount of literature on Fig. 5. Aspect graph [27].
these particular techniques. Vranic et al. [88] presented a fea-
ture representation based on 2-D boundary information, after
having projected the 3-D model onto three standard coordinate
planes, i.e., XY, XZ, and YZ planes. For each projection on a
specified plane, a silhouette was acquired by selecting contour
points, equidistantly or equiangluarly spanned; then, the Fourier
power spectrum was computed. The first n coefficients of the
power spectrum were finally extracted as the feature. The draw-
Fig. 6. Depth image [50].
back is the incapability of properly reflecting the 3-D spatial
information, since the 3-D model was only viewed as a sim-
ple combination of three standard 2-D projections, losing too a set of uniformly distributed cameras by borrowing the concept
much structure information. To solve this problem, Vranic et al. of light field from image-based rendering. The cameras were
added depth information, which encoded the spatial distance controlled to rotate many times when measuring the similarity
difference of 3-D surfaces into different gray values of their between descriptors of two 3-D models, so as to be switched
2-D projection images [88]. Also, they replaced contour-based onto their different vertices. The final 3-D model retrieval results
2-D shape matching with region-based, which also increased the were combined from the matching results of all those acquired
retrieval precision. Cyr et al. [27] proposed an aspect-graph 2-D images by integrating 2-D Zernike moment and Fourier de-
approach to represent 3-D shapes, as shown in Fig. 5. First, scriptors. Ohbuchi et al. [50] presented a similar method. They
2-D projection views were computed according to the view an- generated a depth or z-value image of a 3-D model from mul-
gles achieved after partitioning the viewing sphere by every 5 . tiple viewpoints v that were equally spaced on the unit sphere.
Then, similar 2-D projection views were clustered into the same The 3-D model matching was then performed by adopting a 2-D
group so as to generate a number of clusters called Aspect, Fourier descriptor [89] for similarity matching of 2-D images.
from which the shape representation was created by selecting a The main difference is that Chens 2-D image only contains sil-
representative view for each Aspect. Min et al. [3] projected houettes while Ohbuchis has depth information. Fig. 6 depicts
each 3-D model into several 2-D silhouette images from m dif- Ohbuchis feature extraction process. The depth image was first
ferent viewpoints and then matched all their combinations with mapped from the Cartesian coordinate into the polar coordinate
n (m > n) 2-D sketches drawn by the user or the counterpart to perform Fourier transformation before Fourier descriptors
combinations of other 3-D models. The similarity was mea- were computed.
sured as the minimal sum of all the pairwise sketch-to-image Since many more features can be extracted for a 2-D shape,
(or, image-to-image) similar scores. Chen et al. [24] proposed the function mapping methods make the retrieval process more
a light field descriptor representing the 4-D light field of a 3-D flexible. They can also largely reduce the complexity of feature
model with a collection of 2-D images, which were captured by computation and make the feature descriptor more compact.
1088 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICSPART C: APPLICATIONS AND REVIEWS, VOL. 37, NO. 6, NOVEMBER 2007
However, this inevitably causes much loss of important 3-D
information, since the function mapping process is restricted by
different constraints. Moreover, for 2-D planar view mapping,
how to decide the necessary number of 2-D projection views is
another problem in practice [27].
3) Statistical Properties Methods:
a) Moment: Shapes can also be compared based on their
statistical properties. The most extensively used statistical prop-
erty is the moments, such as Hus image moments [90]. For shapes, Novotni et al. [96] introduced the so-called distance
example, Elad et al. [91] assumed the 3-D model to be a hollow
model bounded by its surfaces, and computed 3-D moment of
surfaces as the 3-D shape representation by
 of the other one, and the average distance from the boundary
mpq r = xp y q z r dxdydz (3)
D
where D is the 3-D model, D is the surface of D, and mpq r is
the (p,q,r)th 3-D moment. For a 3-D model, the set of moments
mpq r is unique so that it constitutes a full and complete descrip-
tion of D, and a partial object description can also be obtained
by using some subset of these moments [92].
However, the main drawback of these methods is that a unit-
scale coordinate frame of the 3-D model has to be acquired prior
to the feature computation process. To address this issue, some
new statistical feature extraction approaches with no need of
pose registration have been proposed. Shape feature based on
3-D Zernike moments [93] is an example. Novotni et al. [63]
demostrated that 3-D Zernike moments are computed as a pro-
jection from the function defining the 3-D object onto a set of
orthonormal functions within a unit sphere, which have sim-
ple representation but good retrieval performance. They further
presented 3-D Zernike invariants as the 3-D shape descriptor.
A 3-D model was transformed into a voxel grid representation
of specific precision and then scaled into a unit sphere, so the
geometrical moments were computed as follows:
 a group of -values spaced at power of two intervals. -shapes
mpq r = f (x, y, z)xp y q z r dxdydz. (4)
|x 2 +y 2 +z 2 |1
Finally, the 3-D Zernike invariants were extracted on the ba-
sis of those computed geometrical moments. The 3-D Zernike
invariants were reported to gain robustness against topological
and geometrical deformations.
b) Histogram: There are also many other kinds of statisti-
cal property features expressed in the form of different discrete
histograms of geometrical statistics [94]. The shape representa-
tion is simplified as a probability distribution problem by using
histograms and avoids the model normalization process. In lit-
erature [95], 3-D histograms were constructed on the crease an-
gles for all edges in a 3-D triangular mesh to match 3-D shapes.
Paquet et al. [16] presented histogram features, including color
histogram, normal vector histogram, and material histogram to
represent 3-D shapes. Paquet et al. also pointed out that a his-
togram can represent the 3-D data distributions, based on voxels,
and is transformation invariant. Ankerst et al. [34] subdivided
the space into shells and sectors around the centroid of a 3-D
object, making the resulting partitions correspond to the bins
of a 3-D shape histogram. In the MPEG-7 standard, there is
also a shape histogram descriptor for 3-D mesh model known as
Fig. 7. Shape functions [2].
the 3-D shape spectrum descriptor (3-DSSD) [5]. For rigid 3-D
histograms as a basic representation. Their fundamental idea
was that if two objects were similar, only a small part of the
volume of one of the objects would be outside the boundary
would also be small. They first computed the offset hulls of each
object based on a 3-D distance field, and then constructed the
distance histograms for each object to indicate how much of the
volume of one object was inside the offset hull of the other.
A specific set of shape functions can also be defined for
histogram computation purposes in order to integrate probability
distributions of geometrical properties. For instance, Osada et al.
[2] computed different histograms, based on several self-defined
shape functions known as shape distribution (including angle,
distance, area, and volume, as shown in Fig. 7), which were all
calculated from a randomly sampled point set on a 3-D models
surface. Song et al. [97] also adopted a histogram representation,
based on shape functions to match 3-D shape by generating
histograms using the discrete Gaussian curvature and discrete
mean curvature of every vertex of a 3-D triangle mesh.
The introduction of geometrical properties into the histogram,
makes multiresolution shape representation possible. Ohbuchi
et al. [51] proposed a multiresolution shape descriptor, repre-
sented in the form of an ordered set of histograms. They first
defined a multiresolution representation (MRR) feature, speci-
fied as a set of 3-D -shapes [98], which was defined by using
are a generalization of the convex hull of a point set, which
shrinks by gradually developing cavities utill it is identical to
the convex hull when = [98]. Next, a 2-D histogram was
generated for each MRR so that an ordered set of histograms
could be produced as the shape descriptor.
To sum up, many statistical shape feature descriptors are sim-
ple to compute and useful for keeping invariant properties. In
many cases, they are also robust against noise, or the small
cracks and holes that exist in a 3-D model. Unfortunately, as
an inherent drawback of a histogram representation, they pro-
vide only limited discrimination between objects: they neither
preserve nor construct spatial information. Thus, they are often
not discriminating enough to make small differences between
dissimilar 3-D shapes, and usually fail to distinguish different
shapes having the same histogram.
4) Topology Methods: Topology is a relatively high-level
representation. It describes the organization and spatial arrange-
ment information: how vertices are connected to compose sur-
faces with edges. A well-designed graph data structure and graph
algorithm can be adopted to represent the topology and skeleton
characteristic of 3-D models. Therefore, this type of method
usually produces a graph-like structure, rather than numeric
YANG et al.: CONTENT-BASED 3-D MODEL RETRIEVAL: A SURVEY
Fig. 8. Multiresolutional Reeb graph [101].
feature descriptors. Examples are as follows. Bardinet et al. [99]
presented a structured 3-D shape representation based on 3-D
skeleton and medial axes, as an extension for the concept of
2-D medial axis transform (MAT) [100]. First, adequate at-
tributed relational graphs (ARGs), consisting of a set of nodes
with attributes and a set of links, were first generated and the
topological features were then extracted from the node and link
structures of those graphs. Hilaga et al. [101] represented the
topology of a 3-D object as a Reeb graph using a function
of geodesic distance [102] between points on the mesh. The
Reeb graph is a skeleton representation using a continuous scalar
function defined on an object with arbitrary dimensions [103].
An appropriately defined, continuous function is invariant to
affine transformations. Hilaga et al. proposed their topological
feature as a Multi-resolution Reeb Graph (MRG) (as shown
in Fig. 8), using a continuous function based on the distribution
of the geodesic distance, which is defined as follows:
 set construction and a gradient descent algorithm was employed
(v) = g(v, p)dS (5)
pS
where v is a point on a surface S, and g(v, p) represents the
geodesic distance between v and another point p on S, which
is the length of the shortest path from v to p. To produce the
scaling invariance, a normalization step is also used, represented
as follows:
(v) minpS (p) drical surface, a cone surface or a planar surface. Next, the
n (v) = (6)
maxpS (p)
The MRG feature is invariant to translation and rotation and
robust against changes in topology structure caused by a mesh
simplification or subdivision. In consequence, it is discrimina-
tive on different level-of-detail. However, MRG lacks the ability
to correctly distinguish the corresponding parts of 3-D models.
A directed graph structure was also adopted to represent the
skeleton of a 3-D volumetric model [104], where an edge was
directional according to a principle similar to shock graph [105].
First, a volumetric cube was thinned into a skeletal-graph, a
line-like sketch composed of the points on the medial axis of
the medial surface planes. Then, a clustering algorithm was
implemented on the thinned voxels to increase the robustness
against small perturbations on the surface and reduce the number
of graph nodes. An undirected acyclic graph was first generated
out of the skeletal points by applying the minimum spanning
tree (MST) algorithm. After that, the directed graph was finally
1089
Fig. 9. Skeleton graph [104].
constructed by directing the edge from a voxel with the higher
distance to the one with lower distance. Here, the distance means
the minimum distance from a voxel to the boundary of the
volumetric object. Fig. 9 gives an example of the skeleton graph
structure.
Topology analysis can also be carried out by decomposing
a 3-D model as a parametric model of a set of simple ele-
mentary regular shapes. The topology is depicted as the spatial
relationships and arrangements of those basic shapes, such as
generalized cylinders [106], deformable regions [107], shock
scaffold [108], and superquadrics [109]. Ma et al. [110] even
presented a practical approach, using a model based on radial
basis functions (RBFs) to extract 3-D skeletons. For a 3-D polyg-
onal object, the vertices were treated as centers for RBF-level
on each vertex to locate the local maxima in the RBF. Finally,
all the connected maxima pairs were handled using the Snake
method, and the final positions of the Snake sequences were
extracted as the skeleton features. Tal et al. [111] first decom-
posed a mesh into elements called watersheds using a Wa-
tershed decomposition algorithm [112], then fit and classified
them into four kinds of basic shapes: a spherical surface a cylin-
shape signature, an attributed decomposition graph, was then
constructed.
The topological and skeletal shape features are attractive for
3-D retrieval because they are able to capture the significant
shape structures of a 3-D object. Meanwhile, they are relatively
high-level and close to humans intuitive perception, which
makes them useful for defining more natural 3-D query
representation. They can also perform part-matching tasks for
containing both local and global structural properties. For some
kinds of topological representations, they are also robust against
the LOD structure of 3-D models due to their multiresolution
properties. However, 3-D models are not always defined well
enough to be easily and naturally decomposed into a canonical
set of features or basic shapes. In addition, the decomposition
process is usually computationally expensive. Moreover, model
decomposition processes are quite noise-sensitive to small
perturbations of the model. Thus, extra effort is, in turn, required
to handle them [2]. Finally, compared with the comparatively
1090 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICSPART C: APPLICATIONS AND REVIEWS, VOL. 37, NO. 6, NOVEMBER 2007
straightforward indexing and similarity matching algorithms
based on numeric feature vectors, the indexing and matching
algorithms of graph-like representations are relatively more
complex and time-consuming due to the necessary graph
searching processes. And, since there is currently no universal
general-purpose graph matching solution, different graph
matching algorithms need to be designed to accommodate
different graph-like representations.
B. Appearance Representations and Extraction
The 3-D models usually possess multimodal feature descrip-
tors. Besides the shape features, the appearance attributes of
3-D models such as material color, color distribution, and tex-
ture, etc., are also an important part of content-based 3-D model
retrieval. In particular, color and texture databases are neces-
sary to render 3-D models. In many practical applications, 3-D
appearance features, such as smoothness, roughness, and distri-
bution of light, might also be of interest [20], so that 3-D model
databases may also need to be searched according to the se-
lected appearance properties. Besides, the visual perception of
geometry of the human being is indeed influenced by color, and
separate colors are often analyzed as distinct entities in humans
visual system [16]. However, there are still insufficient research
data on the appearance representation and extraction method-
ologies, compared with the abundant literature of 3-D shape
representation and extraction. This is partly due to the diversity
and complexity of appearance attributes. For example, the distri-
bution and spatial relationship of colors in 2-D images or videos
can be successfully defined and represented, whereas, in 3-D
models, this is not the case. Therefore, the issues of appearance
representation and measurement in 3-D situations, particularly
how to integrate appearance information into the shape descrip-
tor, or how to directly derive feature descriptors from appearance
information that can then be combined with traditional shape de-
scriptors to comprehensively feature 3-D models, and similarity
measurements suitable for these appearance-contained feature
descriptors, are necessary and require intensive study. Although
some shape features also contain partial appearance information
such as color and texture, for example, geometry image [87] and
histograms [16], they are still too superficial to depict the 3-D
appearance attributes properly.
To date, the appearance representations adopted in 3-D model
retrieval are mostly related to surface colors or surface textures.
Paquet et al. [16] presented a color feature extraction method by
separately taking into account the material color and its luminos-
ity: on the one hand, the material color attribute was described
using a color histogram of each component of red, green, blue
(RGB) color space; on the other hand, luminosity attributes
were represented, employing seven different histograms of dif-
fuse reflection coefficients, specular reflection coefficients, and
textures. Suzuki et al. [21] presented a color feature extrac-
tion method from a different perspective. An appearance fea-
ture representation based on material colors was proposed. This
method can retrieve 3-D polygonal models according to colors
by reflecting the users preferences from some material color
databases. It is believed that material colors greatly influence
the appearance of 3-D models. In a simple rendering model,
material colors of a 3-D model can be specified by ambient
color, diffuse color, specular color, emissive color, shininess,
and transparency. Each material color item contains several
light values. Since the shading model is given by equations
with these light parameters, switching these values can generate
a large number of different colors and change the appearance
of objects. Hence, Suzuki et al. proposed a color extraction and
matching method to handle the material color databases effi-
ciently, based on the users subjective evaluation scales. First,
users were asked to evaluate and describe material colors for
some portions of the database as a study dataset. User inputs
were then analyzed and a multidimensional space was created,
which reflected the users personalized evaluations of mate-
rial colors. To create a complete, personalized search space, a
set of nonstudied data was mapped into the multidimensional
space. Since the light characteristics of each material color were
known, coordinates of each material color could be predicted,
using multiple-regression analysis. In that way, each material
color could be represented and matched. In addition, Suziki
et al. also evaluated another appearance feature representation
using the surface textures of 3-D models where the higher order
local autocorrelation (HLAC) masks were extracted as texture
features [113].
V. SIMILARITY MATCH
After the feature extraction process, appropriate similarity
measurements should be assigned to measure the content sim-
ilarity. The ideal goal of similarity measurement is twofold: 1)
to make the feature vectors of similar 3-D models as close as
possible in the feature space, and 2) to maintain the largest pos-
sible distances for dissimilar 3-D models. Hence, the mission
of the similarity match is to compute the suitable distances or
dissimilarities in multidimensional feature space between the
user query and all the 3-D models in the database and rank them
in descending order of similarities as well. A variable number
of models are then retrieved by listing the top-ranking items.
At present, the available similarity match methods in content-
based 3-D model retrieval can be categorized into four classes:
1) distance metrics; 2) graph matching; 3) machine learning;
and 4) semantic measures.
A. Distance Metrics
Currently, distance metrics are perhaps the most popular and
widely used similarity matching methods, with many types al-
ready, either proposed or in use for content-based 2-D media
retrieval [6], [114]. A distance metric is a dissimilarity mea-
surement having some particular properties, for which there
is a comprehensive body of research [52], [115][117]. For
content-based 3-D model retrieval, the successfully applied dis-
tance metrics include Manhattan distance [58], [102], Euclidean
distance [3], [63], [118], and Hausdorff distance [66]. The Man-
hattan and Euclidean measurements are both based on Lp dis-
tance (p = 1, 2), that is, Minkowski distance. The Lp distance
between two points in N-dimension space x, y RN is defined
YANG et al.: CONTENT-BASED 3-D MODEL RETRIEVAL: A SURVEY
as


Lp (x, y) = [ N
i=1 (xi yi ) ]
p 1/p
. (7)
All distances are metrics when p 1. Lp distance itself can
also be directly used as a similarity measurement. For example,
Osada et al. [2] employed it to implement a similarity match on
the probability density function of shape distribution features. In
particular, to assign different impacts to different features or to
allow relevance feedback, Euclidean distance is often modified
to weighted Euclidean distance with the weight matrix [2], [61],
[91]. The Hausdorff distance, another frequently used metric,
is defined for comparing two point sets of different sizes as
follows:
h(A, B) = minaA maxbB d(A, B) (8)
where d(A, B) is a distance metric, such as Euclidean distance.
However, it is very sensitive to noise since even a single out-
lier can change the distance [116]. Many other distance metrics
have also been studied for 3-D model retrieval task. Ohbuchi
et al. [53], [58] introduced an elastic-matching distance in or-
der to compensate the larger-than-wanted effect caused by
rigid distance metrics, such as Euclidean distance, and the
results were promising. Tangelder et al. [69] used an improved
earthmovers distance [119] as the distance measure. Some 2-D
distance metrics based on Haar wavelet have also been pro-
posed [42], [120].
B. Graph Matching Algorithms
When two 3-D objects that need comparing are represented by
graph-like structures, specially designed graph matching algo-
rithms are needed in order to perform similarity match. However,
matching two graphs is typically considered as the largest iso-
morphic subgraph problem, which is almost impossible to solve.
Hence, the currently available 3-D shape similarity measure-
ments based on graph matching are all customized to the given
3-D topological features and do not provide general solutions.
For example, Bardinet et al. [99] achieved this by finding
the association matrix M so that an objective function involving
all types of nodes, links, and attributes in the graph was mini-
mized. Some heuristic constraints were also used in the objective
function to guarantee the correctness of graph matching. Hilaga
et al. [101] attached each graph node with several attributes and
defined the similarity between two nodes as the similarity be-
tween their attributes. Then, the similarity for a given set of node
pairs was computed as a whole similarity measure. In [71], [72],
and [105], a graph matching algorithm for 2-D shock graphs was
given (Fig. 10). A structural signature was defined for each
graph node, which characterized the nodes underlying subgraph
structure, whose components were based on the eigenvalues of
the subgraphs adjacency matrix. All the edges in the graph
were discarded and the problem was redefined as finding the
maximum cardinality and minimum weight matching in a bidi-
rectional graph. However, this method cannot be guaranteed to
conform to the hierarchical structure of two graphs. To solve
this problem, this process was combined with a recursive depth-
first search in order to use matching at higher levels to constrain
1091
Fig. 10. Graph matching [105].
matching at lower levels [104]. The graph matching algorithm
usually outputs a number of parameters that can be used to
determine the goodness of the similarity match, such as the
number of nodes matched, and information about which nodes
were matched to other nodes. Further, a coarse-to-fine graph
matching strategy can also be easily adopted.
C. Machine Learning Methods
The basic idea of similarity match based on machine learning
is to train a specific learning classifier to compute and rank simi-
larity degrees on a preselected training sample set with a specific
scale by using machine learning methods such as artificial neural
networks, support vector machines (SVMs), etc. This is particu-
larly appropriate in cases where no suitable distance metric can
effectively measure the similarity, such as, for feature vectors
whose dimensions are too high. In those cases, some appropriate
similarity measurements can be approached by learning hidden
correlations and mappings from result-known training samples,
which allows for great flexibility in the retrieval process. For
example, Ibato et al. [28] utilized an SVM learning algorithm to
obtain similarity degrees of pairs of feature vectors. The SVM is
a binary classifier that creates and uses a nonlinear hyperplane
with maximum margins to the training samples of the given pair
of classes [121]. Ibato et al. carried out many experiments by
combining the transform-invariant D2 shape feature [2] with
the SVM, feeding the feature vector to an SVM to compute
the dissimilarity. Pedro et al. [122] adopted another machine
learning method, self-organized map neural network (SOM) to
implement similarity match based on spin image shape features.
In [123], a weighted similarity function for CAD model classifi-
cation, based on an underlying shape distribution feature repre-
sentation and a K-nearest-neighbor learning algorithm (KNN),
was proposed. Given a set of CAD solid models and correspond-
ing categories, the KNN learning method was used to extract
the related patterns to automatically construct a model classifier
and identify new or hidden classifications using the shape dis-
tribution feature, learning from the stored, correctly categorized
training examples. In addition, probabilistic approaches, such as
Bayes theorem, are also a practical way for similarity matching,
in which specific probabilities on features are calculated and the
3-D model having the highest probability will be identified as
the closest matching result [124].
Furthermore, machine learning methods can also be used to
implement user relevance feedback mechanism in 3-D model
retrieval, iteratively refining the retrieval results step by step, by
1092 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICSPART C: APPLICATIONS AND REVIEWS, VOL. 37, NO. 6, NOVEMBER 2007
making designed reactions to the users interactive evaluations.
This can also achieve a personalized retrieval, based on differ-
ent users preferences. A good example is Elad et als work
on relevance feedback [91], [125]. They made use of the SVM
learning algorithm to derive the optimal weight combination for
a weighted Euclidean distance metric, and made stepwise im-
provements to the similarity match, according to every iteration
of the users interactive evaluation.
D. Semantic Measurements
As the 3-D model retrieval results achieved by low-level fea-
tures have proven not to be as discriminative as people had
expected, this raises another important issue, that is, subjec-
tive semantic measurement in similarity comparison. Moreover,
whether a retrieved 3-D model is relevant and irrelevant to
the query is also judged by the users according to their subjective
perception, related to the semantic content [6]. Consequently,
it is highly important to develop semantic similarity-matching
methods that consider the humans perception in content-based
3-D model retrieval process.
Many approaches that have been proposed in 2-D media re-
trieval to reduce the semantic gap manage to perform simi-
larity measurement based on high-level semantics [6], [7]. One
method is to learn the connections between a 3-D model and a set
of semantic descriptors, or the semantic meanings from those
automatically extracted 3-D model features. This approach is
usually based on machine learning and statistical classification,
which groups 3-D models into semantically meaningful cat-
egories using low-level features so that semantically-adaptive
searching methods can be applied to different categories. Ex-
amples are as follows. Suzuki et al. [66], [126] constructed a
multidimensional scaling mechanism so that semantic keyword
descriptors used in the query and the shape features computed
from the 3-D shapes were strongly correlated, based on a train-
ing data set. The multidimensional scaling mechanism can an-
alyze matrices of similar or dissimilar data by representing the
rows and the columns as a point in a Euclidean space and then
measure their similarities using Euclidean distance. They then
created a special user preference space according to this prin-
ciple, in which a function mapping from 3-D model space was
constructed to integrate semantic keywords and 3-D shapes as
a representation of human subjective perception. Zhang et al.
[127] introduced the concept of hidden annotation to con-
struct a semantic tree of the whole 3-D model database. They
used an active learning method to compute a list of probabilities
for each 3-D model, which indicated the models probability
of having a certain semantic attribute. The list of probabilities
was then utilized to calculate the semantic distance between
two models, or between the user query and a model in the
database. The overall dissimilarity between two models was
finally determined by combining the weighted sum of the se-
mantic distance with the low-level feature distance. In [28], a
novel semantic measurement that could simulate human visual
perception was also presented. It was achieved by employing a
well-trained SVM learning classifier constructed by performing
SVM learning from the tagged similar and dissimilar models in
the retrieval results of the current querying step. An SVM-based
semantic clustering and retrieval method was also successfully
implemented in the prototypical 3-D engineering shape search
system (3-DESS) designed by Purdue University [128].
In addition, some concept hierarchies, such as predefined
domain ontology, can also be introduced into the semantic
measuring process. There is some work involved in building
a fundamental framework for representing and measuring the
semantic information of 3-D models, such as the Aim@shape
project (http://www.aim-at-shape.net) launched by the Euro-
pean Commission in order to implement semantically capable
digital representations of 3-D shapes that are expected to ac-
quire, build, transmit, retrieve and process shapes with their
associated knowledge. This project is an attempt to formalize
shape knowledge (in particular, metadata, used for knowledge-
based shape modeling) and define shape ontologies in specific
contexts used for linking semantic keywords to shape features.
The shape knowledge representation is built on three basic lev-
els: geometric, structural, and semantic, where, at the semantic
level, the association of specific semantics to structured and ge-
ometric models is established through automatic annotation of
shapes or shape parts according to the concepts formalized by
the domain ontology. Furthermore, by introducing a common
formalization framework, it is also possible to build a shared
semantic conceptualization of a multilayered architecture for
shape models.
Another effective method is to perform user relevance feed-
back after each search iteration in the database [91], [129], [130].
This is effective in narrowing the gap between low-level feature
similarity and high-level semantic similarity [91], by which,
what the user has in mind is able to be better captured. To some
extent, it is also seen as a method of semantic measurement and
has been extensively employed in 2-D media retrieval [130],
[131]. In the case of 3-D retrieval, Leifman et al. [132] proposed
a relevance feedback method combining query refinement and
supervised feature extraction at each step, which tried to find
an optimal linear transformation that reweights the low-level
feature components in order to achieve the maximal separation
of the original result set. They found this projection by maxi-
mizing a cost function defined as Fishers Linear Discriminant
Criterion [92]. Atmosukarto et al. [133] also presented a sub-
jective similarity measurement relevance feedback process by
combining various distances measured by different feature rep-
resentations. This was implemented by computing the integer
rank rk (Oi |Oj ) of the 3-D object Oi with respect to 3-D object
Oj using a probability estimation method in the feature space
of the relevant and irrelevant result sets.
VI. QUERY REPRESENTATION AND USER INTERFACE
A content-based 3-D model retrieval system is expected to
allow users to submit their query in a natural and interactive way.
Due to the abundance of content descriptions of 3-D models,
there should be a variety of query specifications to be supported,
including the follows.
1) Query-By-Example (QBE): A 3-D model example is di-
rectly provided as a query. The example model can be a new one
YANG et al.: CONTENT-BASED 3-D MODEL RETRIEVAL: A SURVEY
Fig. 11. Query by 2-D sketch [134].
Fig. 12. Relevance feedback interface [132].
uploaded by users, one of the result models in specific retrieval
process, or anyone in the available databases [3], [16], [118].
2) 2-D Projections: Representing a query with a set of
2-D projection images of a 3-D example model from different
viewpoints [3].
3) 2-D Sketches: Using 2-D shapes sketched interactively by
users. For instance, Min et al. [3], [134] designed an interactive
2-D sketch online interface, by which a query was represented
as n 2-D sketches drawn by the users from several different
viewpoints, as shown in Fig. 11.
4) 3-D Sketches: Using a 3-D shape sketched interactively
by users. Min et al. [23] also implemented a 3-D sketch online
interface based on a 3-D sketch tool Teddy, which was designed
by Igarashi et al. [14], [135].
5) Text: Query interface based on text keywords [3] and/or
semantic descriptions [127].
6) Multimodal Queries: Combinations of multiple query
representations mentioned earlier.
In general, a query that is simultaneously done by integrating
multiple-query specifications is more likely to produce better
results than using any individual one [14], [136]. Moreover, the
user interface of 3-D model retrieval is responsible for display-
ing retrieval results to users in a visual and interactive way as
well, in order to make the users browse them or pursue the next
retrieval iteration easily. Some 3-D model retrieval systems also
introduced an interactive user relevance feedback mechanism
into their query interface. For example, Elad et al. [91] presented
an user relevance feedback interface by providing users a chance
to mark a subset of the initial retrieval results as relevant or
irrelevant (using a tick or a cross symbol, as shown in
Fig. 12). Further, Zhang et al. [76] extended this kind of feed-
back interface by adding a way to mark the extent of relevant
and irrelevant, providing both qualitative and quantitative ad-
justments. Similar work can also be found in Ref. [132] and
Ref. [133]. The iterative refinement can automatically narrow
1093
the perception gap between the retrieval system and the users,
which is expected to enhance the retrieval performance.
VII. PERFORMANCE EVALUATION
To compare and evaluate the effectiveness of the search us-
ing the 3-D model retrieval algorithm, or how well the system
meets the users requirements, retrieval performance evaluation
is essential in content-based 3-D model retrieval.
Since there are many types of specialized 3-D models in dif-
ferent domains, the relevant research work, including versatile
shape representations and similarity measures, may also have
varied results on the retrieval task. As a result, when consid-
ering the performance evaluation issue, the first action will be
to define a relatively common and general-purpose 3-D model
collection as a benchmark database, in order to define a common
method to provide relevance judgments.
A. Three-Dimension Model Benchmark Databases
There are currently several representative 3-D model
databases for performance evaluation purposes, among which,
the Princeton shape benchmark (PSB) [137] is perhaps the most
popular and well-organized. PSB is a publicly available 3-D
model benchmark database containing 1814 classified 3-D mod-
els, which have been collected from the Internet and organized
into hierarchical semantic classifications made by experts. PSB
is provided as separate training and test sets, and each 3-D model
has a set of annotations. The authors have also issued a bundle
of software tools for database analysis and browsing.
In addition to PSB, some other 3-D model databases can also
be employed as standard benchmarks containing a wide variety
of 3-D objects, which have been independently gathered by dif-
ferent research groups. These include the Utrecht databases [69],
MPEG-7 databases [29], and Taiwan databases [138]. There are
also several benchmark databases on specific domains, for ex-
ample, CAD models [12] and 3-D protein structures [36]. More
detailed statistics for most of the currently available 3-D model
databases can be found in [137].
However, since the majority of these 3-D model databases
primarily focus on 3-D shapes, there are currently no standard
benchmark databases for appearance attributes, such as color,
texture, and light distribution. Although PSB can partially per-
form this function, it is still neither ideal nor optimal.
B. Performance Evaluation Methods
The most common evaluation measures used in 3-D model
retrieval are precision and recall, which are introduced from the
information retrieval (IR) community and are widely adopted in
2-D media retrieval [139]. They are defined as:
Precision = Nc /Nr Recall = Nc /Ns (9)
where Nc is the number of retrieved models similar to the query,
Nr is the total number of retrieved models from the query, and
Ns is the number of similar models in the whole database.
They are also usually presented as a Precision-Recall (P-R)
graph [140], which shows how precision falls as more and more
1094 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICSPART C: APPLICATIONS AND REVIEWS, VOL. 37, NO. 6, NOVEMBER 2007

objects are retrieved. A retrieval performance evaluation can be TABLE III

TYPICAL CONTENT-BASED 3-D MODEL RETRIEVAL SYSTEMS
achieved based on the P-R graphs. The closer a precision is to 1,
the better performance is obtained. Moreover, the performance
can also be evaluated in some other aspects in terms of the P-R
graph, such as effectiveness and robustness [141]. However,
since relevant and irrelevant are both judged subjectively
by users, this evaluation is naturally born with subjectivity.
There are also other performance evaluation methods used
for 3-D model retrieval. For example, a similarity matrix mea-
surement was presented as a graphical performance evaluation,
by which a matrix with higher contrast was usually rated either
very similar or very dissimilar according to specifically designed
criteria [23]. Many other types of evaluation measures, such as
best matches, distance image, and tier image have also
been proposed [137].

VIII. EXAMPLES OF CONTENT-BASED

3-D MODEL RETRIEVAL SYSTEMS
For content-based 3-D model retrieval, there are already a
number of prototypes, standalone systems, and Internet-based
search engines implemented and publicized for the purpose of
research. Representative examples are listed in Table III.
Moreover, there also some professional 3-D model retrieval
systems. For example, Ankerst et al. [36], [55] developed
a content-based retrieval system for 3-D protein databases,
while HeriotWatt University implemented a Web-based search
engine, ShapeSifter (URL: http://www.shapesearch.net), and
Drexel University (URL: http://edge.mcs.drexel.edu/repository/
frameset.html) built a digital library for 3-D CAD models and
3-D engineering designs [15], [40]. Another noticeable trend
is the 3-D model retrieval services for handsets such as mo-
bile phones and personal digital assistants. For example, Suzuki
et al. [67] developed a 3-D model retrieval prototype for mobile
phone users. Moreover, a 3-D model retrieval system adopting
the MPEG-7 mechanism can also be easily tailored for Pocket-
PC users [24], [138].

IX. CONCLUSION AND FUTURE RESEARCH

Since its inception, content-based 3-D model retrieval has
witnessed abundant great development and achievements in both
theory and application. Nevertheless, the accuracy of content
information in 3-D models, as a consequence of its versatile
aspects and subjective cognition, is very much in question. Much
work still needs to be undertaken to remedy this situation. The
following are just some of the crucial issues and challenges
deserving further investigation.
1) Research on a unified 3-D model retrieval framework. The
3-D data formats are very diverse while the content of a 3-D
model remains independent of them and, so far, has been our
main focus of attention. A practical unified 3-D retrieval frame-
work is, therefore, needed; one that is capable of accommo-
dating most 3-D data representations adaptively by extract-
ing representation-independent features or performing standard
transformations on-the-fly. Moreover, considering the effi-
ciency of transmitting and retrieving 3-D models on the In-
YANG et al.: CONTENT-BASED 3-D MODEL RETRIEVAL: A SURVEY
ternet, feature extraction and similarity matching directly from
compressed 3-D data are also meaningful.
2) Further development of more discriminative 3-D shape
features, especially those that are normalization-free and possess
strong discriminative power. These must also be natural and
simple for effective index mechanisms.
3) Local partial shape feature extraction, i.e., to implement the
feature vectors that are suitable for partial matching inside a 3-D
model. In practice, partial shape features that can describe the
local details are often needed for more precise multiresolution
and flexible retrieval.
4) Multiple features need to be combined for effective sim-
ilarity search. Some work has already been undertaken toward
this aim [53], [96], [136], [144]. However, when a large num-
ber of feature descriptors are used for the query, the system
may not be able to respond quickly because more computations
are needed. Therefore, feature descriptor selection or reduction
techniques must be designed and applied. Consequently, how to
select and weigh those feature descriptors is also an important
and promising future direction.
5) Further development of nonshape descriptors of 3-D mod-
els, such as material color and texture. Furthermore, extraction
of high-level semantic features and similarity measurements
combining with semantic information will also offer important
research issues and challenges.
6) Research on the mechanism of relevance feedback and
personalized retrieval integrating user preferences, by which
users are able to tune the search criteria by themselves toward
more satisfactory search results.
7) Development of simple but powerful query interfaces. The
3-D sketching tool currently used for 3-D shape queries is not
user-friendly for the novice. A less complex way for users to
build simple 3-D objects and 3-D sketches should be provided,
for example, an interface that allows users to form a complicated
3-D object by connecting some basic shapes, just like using
building blocks [14]. On the other hand, a more effective query
interface that is able to locate objects under a nonrigid-body
transformation should also be designed.
8) Retrieval issues targeted 3-D scenes that contain multiple
3-D models. Currently, retrieval methods are mostly limited to
the single 3-D model. However, in many applications, such as a
virtual reality environment, 3-D models are usually presented in
complex 3-D scenes. Therefore, 3-D model retrieval technology
should be extended to handle more complex 3-D scenes [26],
[145]. A novel hierarchical object structure of 3-D scenes may
need to be investigated, to localize and recognize the 3-D objects
in a 3-D scene.
ACKNOWLEDGMENT
The authors would like to acknowledge the insights and sug-
gestions provided by Dr. M. Barlow and all anonymous review-
ers on this manuscript.
REFERENCES
[1] P. Schroder and W. Sweldens, Digital Geometry Processing. Frontiers
of Engineering Series. Washington, DC: Nat. Acad. Eng. Press, 2001,
pp. 4144.
1095
[2] R. Osada, T. Funkhouser, B. Chazelle, and D. Dobkin, Matching 3D
models with shape distributions, presented at the Int. Conf. Shape
Model. Appl., Genova, Italy, 2001.
[3] P. Min, A. Halderman, M. Kazhdan, and A. Funkhouser, Early expe-
riences with a 3D model search engine, in Proc. Web3D Symp., Saint
Malo, France, 2003, pp. 718.
[4] T. Funkhouser, M. Kazhdan, P. Shilane, P. Min, W. Kiefer, A. Tal,
S. Rusinkiewicz, and D. Dobkin, Modeling by example, ACM Trans.
Graph., vol. 23, no. 3, pp. 652663, Aug. 2004.
[5] MPEG-7 Visual Part of eXperimentation Model, version 9.0 ed., MPEG
Video Group, Pisa, Italy, Jan. 2001.
[6] Y. B. Yang, Research and Applications on the Key Techniques of Content-
Based Image Retrieval. Nanjing, China: Dept. Comput. Sci., Nanjing
Univ., Jun. 2003. (in Chinese with English abstract).
[7] Y. X. Chen and J. Z. Wang, Eds., Machine Learning and Statistical
Modeling Approaches to Image Retrieval. Norwell, MA: Kluwer,
2004.
[8] 3D cafe free 3D models meshes. (2003). [Online]. Available:
http://www.3dcafe. com.
[9] Avalon 3D archive. (2003). [Online]. Available: http://avalon. view-
point.com.
[10] CADLIB Web based CAD parts library. (2003). [Online]. Available:
http://www.cadlib. co.uk.
[11] Meshnose, the 3D objects search engine. (2003). [Online]. Available:
http://www.deepfx.com/meshnose.
[12] National design repository. (2003). [Online]. Available: http://www.
deepfx.com/meshnose.
[13] J. W. Tangelder and R. C. Veltkamp, A survey of content based 3D
shape retrieval methods, in Proc. Shape Model. Int., 2004, pp. 145156.
[14] T. Funkhouser, P. Min, and M. Kazhdan et al., A search engine for 3D
models, ACM Trans. Graph., vol. 22, pp. 83105, Jan. 2003.
[15] J. Corney, H. Rea, and D. Clark et al., Coarse filters for shape match-
ing, IEEE Comput. Graph. Appl., vol. 22, no. 3, pp. 6574, May/Jun.
2002.
[16] E. Paquet and M. Rioux, A content-based search engine for VRML
databases, in Proc. IEEE Int. Conf. Comput. Vis. and Pattern Recognit.,
Santa Barbara, CA, USA, 1998, pp. 541546.
[17] W. Regli and V. Cicirello, Managing digital libraries for computer-aided
design, Comput. Aided Des., vol. 32, pp. 110132, 2000.
[18] D. McWherter, M. Peabody, W. Regli, and A. Shokoufandeh, Transfor-
mation invariant shape similarity comparison of solid models, presented
at the ASME DETC. Pittsburgh, PA, Sep. 2001.
[19] S. Mukai, S. Furukawa, and M. Kuroda, An algorithm for deciding
similarities of 3-d objects, presented at the ACM Symp. Solid Model.
Appl., Saarbrucken, Germany, Jun. 2002.
[20] D. V. Vranic and D. Saupe, 3D model retrieval, presented at the
Spring Conf. Comput. Graph. (SCCG 2000). Comenius University Press,
Budmerice Slovakia, May 2000.
[21] M. Suzuki, A web-based retrieval system for 3D polygonal models, in
Proc. Joint 9th IFSA World Congr. 20th NAFIPS (IFSA/NAFIPS 2001),
Vancouver, BC, Canada, Jul. 2001, pp. 22712276.
[22] M. Kazhdan, Shape representations and algorithms for 3D model
retrieval, Ph.D. dissertation, Dept. Comput. Sci., Princeton Univ.,
Princeton, NJ, Jun. 2004.
[23] P. Min, A 3D model search engine, Ph.D. dissertation, Dept. Comput.
Sci. Princeton Univ., Princeton, NJ, Jan. 2004.
[24] D. Y. Chen, Three-dimensional model shape description and retrieval
based on lightfield, Ph.D. dissertation, Dept. Compute. Sci. Inf. Eng.,
National Taiwan Univ., Taipei, Taiwan, Jun. 2003.
[25] T. Seidl, Adaptable similarity search in 3-d spatial database systems,,
Ph.D. dissertation Faculty Math. Comput. Sci., Univ. Munich,. Munich,
Germany, 1997.
[26] A. Johnson and M. Hebert, Using spin images for efficient object recog-
nition in cluttered 3D scenes, IEEE Trans. Pattern Anal. Mach. Intell.,
vol. 21, no. 5, pp. 433449, May 1999.
[27] C. Cyr and B. Kimia, 3D object recognition using shape similarity-based
aspect graph, in Proc. 8th IEEE Int. Conf. Comput. Vision, Vancouver,
BC, Canada, 2001, pp. 254261.
[28] M. Ibato, T. Otagiri, and R. Ohbuchi, Shape-similarity search of three-
dimensional models based on subjective measures, IPSJ SIG Notes
Graph. CAD, vol. 16, pp. 2530, 2002.
[29] T. Zaharia and F. Preteux, 3D versus 2D/3D shape descriptors: A com-
parative study,, in Proc. SPIE Conf. Image Process.: Algorithms Syst.
IIISPIE Symp. Electron. Imaging, Sci. Technol., San Jose, CA, Jan.,
vol. 5298, pp. 4758.
1096 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICSPART C: APPLICATIONS AND REVIEWS, VOL. 37, NO. 6, NOVEMBER 2007
[30] N. Iyer, K. Lou, S. Jayanti, Y. Kalyanaraman, and K. Ramani, Three
dimensional shape searching: State-of-the-art review and future trends,
Comput. Aided Des., vol. 37, no. 5, pp. 509530, 2005.
[31] T. Zaharia and F. Preteux, 3D shape-based retrieval within the MPEG-7
framework, in Proc. SPIE Conf., Jan. 2004, vol. 4304, pp. 133145.
[32] D. V. Vranic and D. Saupe, A feature vector approach for retrieval of
3D objects in the context of mpeg-7, in Proc. Int. Conf. Augmented,
Virtual Environ. Three-Dimensional Imaging (ICAV3D 2001), Mykonos,
Greece, May, pp. 3740.
[33] H. Kriegel and T. Seidl, Approximation-based similarity search for 3-d
surface segments, Geoinformatica, vol. 2, no. 2, pp. 113147, 1998.
[34] M. Ankerst, G. Kastenmuller, H. Kriegel, and T. Seidl, 3D shape his-
tograms for similarity search and classification in spatial databases, in
Proc. Symp. Large Spatial Databases, 1999, pp. 207226.
[35] D. Keim, Efficient geometry-based similarity search of 3D spatial
databases, presented at the ACM SIGMOD Int. Conf. Manage. Data,.
Philadelphia, PA, 1999.
[36] H. Berman, J. Westbrook, Z. Feng, G. Gilliland, T. N. Bhat, H. Weissig,
I. N. Shindyalov, and P. E. Bourne, The protein data bank, Nucleic
Acids Res., vol. 28, pp. 235242, 2000.
[37] G. Kastenmluer, H. Kriegel, and T. Seidl, Similarity search in 3D protein
databases, presented at the 6th German Conf. Bioinformatics, Cologne,
Germany, 1998.
[38] K. Lou, S. Prabhakar, and K. Ramani, Content-based three-dimensional
engineering shape search, in Proc. 20th Int. Conf. Data Eng. (ICDE
2004), Boston, MA, Mar./Apr, pp. 754765.
[39] N. Iyer, Y. Kalyanaraman, K. Lou, S. Jayanti, and K. Ramani, A recon-
figurable, intelligent 3D engineering shape search system. Part I: Shape
representation, presented at the ASME DETC 2003, Comput. Inform.
Eng. (CIE) Conf., Chicago, IL.
[40] D. McWherter, M. Peabody, A. Shokoufandeh, and W. Regli, Solid
model databases: Techniques and empirical results, ASME/ACM Trans.,
J. Comput. Inf. Sci. Eng., vol. 1, no. 4, pp. 300310, Dec. 2001.
[41] 3D knowledge: Acquisition, representation and analysis. (2003). [On-
line]. Available: http://3dk.asu.edu.
[42] W. H. Rynson and W. Ben, Web-based 3D geometry model retrieval,
World Wide Web: Internet Web Inf. Syst., vol. 5, no. 3, pp. 193206,
2002.
[43] J. Loffler, Content-based retrieval of 3D models in distributed web
databases by visual shape information, in Proc. Int. Conf. Inf. Vis., Jul.
2000, p. 82.
[44] M. Suzuki, A search engine for polygonal models to support develop-
ment of 3D e-learning applications, in Proc. 10th Int. World Wide Web
Conf., 2001, pp. 182183.
[45] A. W. M. Smeulders, M. Worring, S. Santini, A. Gupta, and R. Jain,
Content-based image retrieval at the early years, IEEE Trans. Pattern
Anal. Mach. Intell., vol. 22, no. 12, pp. 13491380, Dec. 2000.
[46] D. Vranic and D. Saupe, Description of 3D-shape using a complex
function on the sphere, in Proc. IEEE Int. Conf. Multimedia Expo (ICME
2002), Lausanne, Switzerland, Aug., pp. 177180.
[47] T. Murali and T. Funkhouser, Consistent solid and boundary repre-
sentations from arbitrary polygonal data, in Comput. Graph. (1997
SIGGRAPH Symp. Interact. 3D Graph.), pp. 155162.
[48] A. Gueziec, G. Taubin, F. Lazarus, and W. Horn, Converting sets of
polygons to manifold surfaces by cutting and stitching, in Proc. IEEE
Vis., 1998, pp. 383390.
[49] R. Osada, T. Funkhouser, B. Chazelle, and D. Dobkin, Shape dis-
tributions, ACM Trans. Graph., vol. 21, no. 4, pp. 807832, Oct.
2002.
[50] R. Ohbuchi, M. Nakazawa, and T. Takei, Retrieving 3D shapes based on
their appearance, in Proc. 5th ACM SIGMM, Int. Workshop Multimedia
Inf. Retrieval, Berkeley, CA, 2003, pp. 3945.
[51] R. Ohbuchi and T. Takei, Shape-similarity comparison of 3D models
using alpha shapes, in Proc. 11th Pacific Conf. Comput Graph. Appl.
(PG 2003), Canmore, AB, Canada, Oct., pp. 293302.
[52] T. Hlavaty, 3D object classification and retrieval, Depart. Comput. Sci.
Eng., Univ. West Bohemia. Pilsen, Czech Republic, Mar. 2003, Tech.
Rep. DCSE/TR-2003-05.
[53] R. Ohbuchi, T. Minamitani, and T. Takei, Shape-similarity search of
3D models by using enhanced shape functions, in Proc. Theory Pract.
Comput. Graph, Jun. 2003, pp. 97105.
[54] M. Kazhdan, T. Funkhouser, and S. Rusinkiewicz, Shape matching and
anisotropy, ACM Trans. Graph., vol. 23, no. 3, pp. 623629, Aug. 2004.
[55] M. Ankerst, G. Kastenmuller, H. Kriegel, and T. Seidl, Nearest neighbor
classification in 3D protein databases, in Proc. ISMB, 1999, pp. 3443.
[56] K. Fukunaga, Introduction to Statistical Pattern Recognition, 2nd ed.
New York: Academic, 1990.
[57] D. Vranic, D. Saupe, and J. Richter, Tools for 3D-object retrieval:
Karhunen-loeve transform and spherical harmonics, in Proc. IEEE 2001
Workshop Multimedia Signal Process., Cannes, France, Oct., pp. 293
298.
[58] R. Ohbuchi, T. Otagiri, M. Ibato, and T. Takei, Shape-similarity search of
three-dimensional models using parameterized statistics, in Proc. 10th
Pacific Conf. Comput. Graph. Appl., Beijing, China, 2002, pp. 265275.
[59] E. Paquet, A. Murching, T. Naveen, A. Tabatabai, and M. Roux, De-
scription of shape information for 2-D and 3-D objects, Signal Process.:
Image Commun., vol. 16, pp. 103122, 2000.
[60] M. Heczko, D. Keim, D. Saupe, and D. Vranic, A method for similarity
search of 3D objects, (in German) in Proc. BTW 2001,. Oldenburg,
Germany, Mar., pp. 384401.
[61] M. Kazhdan, B. Chazelle, and D. Dobkin et al., A reflective symmetry
descriptor, in Proc. Eur. Conf. Comput. Vision (ECCV), May 2002,
pp. 642656.
[62] M. Kazhdan, T. Funkhouser, and S. Rusinkiewicz, Rotation invariant
spherical harmonic representation of 3D shape descriptors, presented
at the Eurograph./ACM SIGGRAPH Symp. Geom. Process. Aachen,
Germany.
[63] M. Novotni and R. Klein, 3D Zernike descriptors for content based
shape retrieval, presented at the 8th ACM Symp. Solid Model. Appl.
Washington, Washington, DC, 2003.
[64] L. Kolonias, D. Tzovaras, S. Malassiotis, and M. Strintzis, Fast content-
based search of VRML models based on shape descriptors, in Proc.
IEEE Int. Conf. Image Process, Thessaloniki, Greece, Oct. 2001, vol. 2,
pp. 133136.
[65] M. Yu, I. Atmosukarto, W. K. Leow, Z. Huang, and R. Xu, 3D model
retrieval with morphing-based geometric and topological feature maps,
in Proc. IEEE Conf. Comput. Vis. Pattern Recognit, 2003, pp. 656
661.
[66] M. Suzuki, T. Kato, and N. Otsu, A similarity retrieval of 3D polygonal
models using rotation invariant shape descriptors, in Proc. IEEE Int.
Conf. Syst., Man, Cybern. (SMC 2000), Nashville, TN, pp. 29462952.
[67] M. Suzuki, Y. Yaginuma, and Y. Sugimoto, A 3D model retrieval sys-
tem for cellular phones, in Proc. IEEE Int. Conf. Syst Man Cybern,
Washington, DC, Oct. 2003, pp. 38463851.
[68] D. Vranic and D. Saupe, 3D shape descriptor based on 3D fourier
transform, in Proc. EURASIP Conf. Digital Signal Process. Multimedia
Commun. Services (ECMCS 2001), K. Fazekas, Ed., Budapest, Hungary,
Sep., pp. 271274.
[69] J. Tangelder and R. Veltkamp, Polyhedral model retrieval using
weighted point sets, in Int. J. Image Graph., vol. 3, pp. 121,
2003.
[70] V. Cicirello and W. Regli, Machining feature-based comparisons of
mechanical parts, in Proc. Int. Conf. Shape Model. Appl., May 2001,
pp. 176185.
[71] A. Shokoufandeh, S. J. Dickinson, K. Siddiqi, and S. W. Zucker, Index-
ing using a spectral encoding of topological structure, in Proc. Comput.
Vis. Pattern Recognit., 1999, vol. 2, pp. 491497.
[72] A. Shokoufandeh, S. Dickinson, C. Jonsso, L. Bretzner, and T. Lindeberg,
On the representation and matching of qualitative shape at multiple
scales, in Proc. 7th Eur. Conf. Comput. Vis., Copenhagen, Denmark,
May 2002, pp. 759775.
[73] H. Chen and T. Huang, A survey of construction and manipulation of
octrees, Comput. Vis., Graph., Image Process., vol. 43, pp. 409431,
1988.
[74] S. Yang and T. Lin, A new linear octree construction by filling algo-
rithms, in Proc. 10th Annu. Int. Phoenix Conf., 1991, pp. 740746.
[75] Y. Kitamura and F. Kishino, A parallel algorithm for octree genera-
tion from polyhedral shape representation, Pattern Recognit., vol. 3,
pp. 303309, 1996.
[76] C. Zhang and T. Chen, Efficient feature extraction for 2d/3d objects in
mesh representation, presented at the ICIP 2001. Thessaloniki, Greece.
[77] B. Horn, Extended gaussian images, Proc. IEEE, vol. 72, no. 12,
pp. 16711686, Dec. 1984.
[78] H. Delingette, M. Hebert, and K. Ikeuchi, A spherical representation for
the recognition of curved objects, in Proc. ICCV, 1993, pp. 103112.
[79] D. Saupe and D. Vranic, 3D model retrieval with spherical harmonics
and moments, in Proc. DAGM 2001, B. Radig and S. Florczyk, Eds.,
Munich, Germany, Sep., pp. 392397.
[80] C. Zahn and R. Roskies, Fourier descriptors for plane closed curves,
IEEE Trans. Comput., vol. 21, no. 3, pp. 269281, Mar. 1972.
YANG et al.: CONTENT-BASED 3-D MODEL RETRIEVAL: A SURVEY
[81] G. Leifman, S. Katz, A. Tal, and R. Meir, Signatures of 3D models for
retrieval, in Proc. Israel-Korea Bi-Nat. Conf. Geom. Model. Comput.
Graph., Feb. 2003, pp. 159163.
[82] M. Kazhdan, T. Funkhouser, and S. Rusinkiewicz, Symmetry descrip-
tors and 3D shape matching, in Symp. Geometry Process. Jul. 2004,
pp. 117126.
[83] D. Zhang and M. Hebert, Harmonic shape images: A 3D free-form
surface representation and its applications in surface matching, in Proc.
Energy Minimization Methods Comput. Vis. Pattern Recognit., M. Pellilo
and E. Hancock, Eds. New York: Springer Verlag, 1999, pp. 3043.
[84] J. Eells and L. H. Sampson, Harmonic mappings of riemannian mani-
folds, Amer. J. Math., vol. 86, pp. 109160, 1964.
[85] X. Gu, S. Gortler, and H. Hoppe, Geometry images, in Proc. ACM
Siggraph 2002, pp. 355361.
[86] E. Praun and H. Hoppe, Spherical parametrization and remeshing, in
Proc. Siggraph 2003, pp. 340349.
[87] H. Laga, H. Takahashi, and M. Nakajima, Geometry image matching
for similarity estimation of 3D shapes, in Proc. Comput. Graph. Int.,
Crete, Greece, Jun. 2004, pp. 490496.
[88] D. Vranic, 3D model retrieval, Ph.D. dissertation, Univ. Leipzig,
Leipzig, Germany, 2004.
[89] D. S. Zhang and G. Lu, Shape-based image retrieval using generic
fourier descriptor, Signal Process.: Image Commun., vol. 17, no. 10,
pp. 825848, Nov. 2002.
[90] M. K. Hu, Visual pattern recognition by moment invariants, IRE Trans.
Inf. Theory, vol. 8, no. 2, pp. 179187, 1962.
[91] M. Elad, A. Tal, and S. Ar, Content based retrieval of VRML objects
An iterative and interactive approach, in Proc. 6th Eurograph. Workshop
Multimedia, Manchester, U.K., 2001, pp. 97108.
[92] R. M. Duda and P. E. Hart, Pattern Classification and Scene Analysis.
New York: Wiley, 1973.
[93] N. Canterakis, 3-D Zernike moments and Zernike affine invariants for
3D image analysis and recognition, presented at the 11th Scand. Conf.
Image Anal., Kangerlussaq, Greenland, 1999, vol. 3.
[94] A. P. Ashbrook, N. A. Thacker, P. I. Rockett, and C. I. Brown, Robust
recognition of scaled shapes using pairwise geometric histograms, in
Proc. BMVC, Birmingham, U.K., 1995, pp. 503512.
[95] P. Besl, Triangles as a primary representation, object recognition in
computer vision, in Lecture Notes in Computer Science, vol. 1929. New
York: Springer-Verleg, pp. 11911206, 1994.
[96] M. Novotni and R. Klein, A geometric approach to 3D object compari-
son, in Proc. Int. Conf. Shape Model. Appl., Genova, Italy, May 2001,
pp. 167175.
[97] J. J. Song and F. Golshani, Shape-based 3D model retrieval, in Proc.
15th IEEE Int. Conf. Tools Artif. Intell., Nov. 2003, pp. 636640.
[98] H. Edelsbrunner and E. P. Mucke, Three-dimensional alpha shapes,
ACM Trans. Graph., vol. 13, no. 1, pp. 4372, 1994.
[99] E. Bardinet, S. Vidal, S. Arroyo, G. Malandain, and N. Capilla, Struc-
tural object matching, presented at the Adv. Concepts Intell. Vision
Syst. (ACIVS 2000), Baden-Baden, Allemagne, Aug.
[100] H. Blum, Biological shape and visual science, J. Theoret. Biol., vol. 38,
pp. 205287, 1973.
[101] M. Hilaga, Y. Shinagawa, T. Kohmura, and T. Kunii, Topology matching
for fully automatic similarity estimation of 3D shapes, presented at the
SIGGRAPH 2001, Los Angeles, CA.
[102] M. Sharir and A. Schorr, On shortest paths in polyhedral spaces, SIAM
J. Comput., vol. 15, no. 1, pp. 193215, 1986.
[103] G. Reeb, On the singular points of a completely integrable PfAFF form
or of a numerical function, (in French) Comptes Randus Acad. Sci.,
vol. 222, pp. 847849, 1946.
[104] H. Sundar, D. Silver, N. Gagvani, and S. Dickinson, Skeleton based
shape matching and retrieval, in Proc. Shape Model. Int. 2003, Seoul,
Korea, May, pp. 130139.
[105] K. Siddiqi, A. Shokoufandeh, S. Dickinson, and S. Zucker, Shock graphs
and shape matching, Comput. Vis., pp. 222229, 1998.
[106] T. Binford, Visual perception by computer, in Proc. IEEE Conf. Syst.
Sci. Cybern., Miami, FL, 1971.
[107] R. Basri, L. Costa, D. Geiger, and D. Jacobs, Determining the similarity
of deformable shapes, Vis. Res., vol. 38, pp. 23652385.
[108] F. Leymarie and B. Kimia, The shock scaffold for representing 3D
shape, in Proc. 4th Int. Workshop Visual Form, 2001, pp. 216228.
[109] Y. Zhang, A. Koschan, and M. Abidi, Superquadrics based 3D object
representation of automotive parts utilizing part decomposition, in Proc.
SPIE 6th Int. Conf. Qual. Control Artif. Vis., May 2003, vol. 5132,
pp. 241251.
1097
[110] W. Ma, F. Wu, and M. Ouhyoung, Skeleton extraction of 3D objects
with radial basis functions, in Proc. Shape Model. Int. 2003, Seoul,
Korea, May, pp. 207216.
[111] A. Tal and E. Zuckerberger, Mesh retrieval by components, presented
at the Int. Conf. Comput. Graph. Theory Appl., Set ubal, Portugal, Feb.
2006.
[112] J. C. Serra, Image Analysis and Mathematical Morphology, 1st ed.
London, U.K.: Academic, 1982.
[113] M. Suzuki, Y. Yaginuma, and Y. Shimizu, A texture similarity evaluation
method for 3D models, in Proc. Int. Conf. Internet Multimedia Syst.
Appl. (IMSA 2005), pp. 185190.
[114] S. Santini and R. Jain, Similarity measures, IEEE Trans. Pattern Anal.
Mach. Intell., vol. 21, no. 9, pp. 871883, Sep. 1999.
[115] A. Tversky, Features of similarity, Psychol. Rev., vol. 84, no. 4,
pp. 327352, 1977.
[116] R. C. Veltkamp and M. Hagedoorn, Shape similarity measures, prop-
erties and constructions, in Proc. VISUAL 2000, Lyon, France: Lecture
Notes in Computer Science, vol. 1929. New York: Springer-Verlag, Nov.,
pp. 467476.
[117] R. C. Veltkamp, Shape matching: Similarity measures and algorithms,
in Proc. Shape Model. Int., May 2001, pp. 188197.
[118] B. David, Methods for content-based retrieval of 3D models, presented
at the 3rd Annu. CM316 Conf. Multimedia Syst., Southampton, U.K.,
Jan. 2003.
[119] Y. Rubner, C. Tomasi, and L. J. Guibas, The earth movers distance
as a metric for image retrieval, Int. J. Comput. Vis., vol. 40, no. 2,
pp. 99121, Nov. 2000.
[120] J. Gain and J. Scott, Fast polygon mesh querying by example, in Proc.
SIGGRAPH Tech. Sketches, Aug. 1999, p. 241.
[121] V. Vapnik, The Nature of Statistical Learning Theory, 2nd ed. New York:
Springer-Verlag, 1999.
[122] A. Pedro, D. Alberto, and M. Jose, Spin images and neural networks
for efficient content-based retrieval in 3D object databases, in Proc.
CIVR 2002, Lecture Notes in Computer Science, vol. 2383. New York:
Springer-Verleg, pp. 225234.
[123] C. Ip, W. Regli, L. Sieger, and A. Shokoufandeh, Automated learning of
model classifications, in Proc. ACM Symp. Solid Model. Appl. Archive,
Washington, DC, Jun. 2003, pp. 322327.
[124] T. Ansary, J. Vandeborre, S. Mahmoudi, and M. Daoudi, A Bayesian
framework for 3D models retrieval based on characteristic views, in
Proc. 2nd Int. Symp. 3D Data Process., Vis. Transmiss. (3DPVT 2004),
Sep., pp. 139146.
[125] M. Elad, A. Tal, and S. Ar, Directed search in a 3D objects database
using svm, HP Laboratories, Haifa, Israel, Tech. Rep. HPL-2000-20R1,
2000.
[126] M. Suzuki, T. Kato, and H. Tsukune, 3D object retrieval based on
subjective measures, in Proc. 9th Int. Conf. Database Expert Syst. Appl.
Workshop (DEXA), 1998, pp. 850855.
[127] C. Zhang and T. Chen, Active learning for information retrieval: Using
3D models as an example, Carnegie Mellon Univ., Pittsburgh, PA, Tech.
Rep. AMP01-04, 2001.
[128] S. Hou, K. Lou, and K. Ramani, SVM-based semantic clustering and
retrieval of a 3D model database, in J. Comput. Aided Des. Appl., Proc.
CAD 2005, Bangkok, Thailand, Jun., vol. 2, pp. 155164.
[129] Y. Rui, T. S. Huang, M. Ortega, and S. Mehrotra, Relevance feed-
back: A power tool in interactive content-based image retrieval, IEEE
Trans. Circuits Syst. Video Technol., vol. 8, no. 5, pp. 644655,
Sep. 1998.
[130] Y. Ishikawa, R. Subramanya, and C. Faloutsos, Mindreader: Query
databases through multiple examples, presented at the 24th VLDB
Conf., New York, 1998.
[131] Y. Rui, T. S. Huang, and S. F. Chang, Image retrieval: Current tech-
niques, promising directions, and open issues, J. Vis. Commun. Image
Represent., vol. 10, no. 1, pp. 3962, 1999.
[132] G. Leifman, R. Meir, and A. Tal, Relevance feedback for 3D shape
retrieval, in Proc. IsraelKorea Bi-Nat. Conf. Geom. Model. Comput.
Graph., Oct. 2004, pp. 1519.
[133] I. Atmosukarto, W. K. Leow, and Z. Huang, Feature combination and
relevance feedback for 3D model retrieval, in Proc. 11th Int. Multimedia
Model. Conf. (MMM 2005), Melbourne, Australia, Jan., pp. 128133.
[134] P. Min, J. Chen, and T. Funkhouser, A 2D sketch interface for a 3D
model search engine, in Proc. SIGGRAPH Tech. Sketches, 2002, p. 138.
[135] T. Igarashi, S. Matsuoka, and H. Tanaka, Teddy: A sketching interface
for 3D freeform design, in Proc. SIG-GRAPH 1999. Los Angeles, CA:
ACM, pp. 409416.
1098 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICSPART C: APPLICATIONS AND REVIEWS, VOL. 37, NO. 6, NOVEMBER 2007

[136] P. Min, M. Kazhdan, and T. Funkhouser, A comparison of text and Hui Lin (M00) received the Graduate degree in
shape matching for retrieval of online 3D models, presented at the Eur. aerophotogrammetric engineering from the Wuhan
Conf. Digit. Libraries, Bath, U.K., Sep. 2004. Technical University of Surveying and Mapping,
[137] P. Shilane, M. Kazhdan, P. Min, and T. Funkhouser, The Princeton shape Wuhan, China, in 1980, the M.Sc. degree in remote
benchmark, in Proc. Shape Model. Int., Genoa, Italy, 2004. sensing and cartography from the Graduate School
[138] D. Chen, X. Tian, Y. Shen, and M. Ouhyong, On visual similarity based of Chinese Academy of Sciences, Beijing, China, in
3D model retrieval, in Proc. Comput. Graph. Forum, Sep. 2003, vol. 22, 1983, and the Ph.D. degree in geographical informa-
no. 3 pp. 223232. tion systems from the State University of New York
[139] K. Jarvelin and J. Kekalainen, IR evaluation methods for retrieving at Buffalo, Buffalo, in 1992.
highly relevant documents, in Proc. 23rd ACM SIGIR Conf. Res. Dev. He is currently a Professor and the Director of
Inf. Retrieval, 2000, pp. 4148. the Joint Laboratory for GeoInformation Science,
[140] J. Rocchio, Relevance feedback in information retrieval, in The SMART Chinese University of Hong Kong, Shatin, N.T., Hong Kong. His current re-
Retrieval System: Experiments in Automatic Document Processing. En- search interests include virtual geographic environments, cloud-prone and rainy
glewood Cliffs, NJ: Prentice-Hall, 1971, pp. 313323. area remote sensing, spatially integrated humanities, and social science.
[141] B. Bustos, D. Keim, D. Saupe, T. Schreck, and D. Vranic, An experi- Prof. Lin was elected as an Academician of the International Eurasian
mental comparison of feature-based 3D retrieval methods, in Int. Symp. Academy of Sciences in 1995.
3D Data Process., Vis., Transmiss., Thessaloniki, Greece, Sep. 69, 2004,
pp. 215222.
[142] E. Paquet and M. Rioux, Nefertiti: A query by content software for
three-dimensional models databases management, Image Vis. Comput.,
vol. 17, no. 2, pp. 157166, 1999.
[143] M. Novotni and R. Klein, Geometric 3D comparisonAn application,
presented at the ECDL WS Generalized Documents 2001, Darmstadt,
Germany, 2001.
[144] B. Bustos, D. Keim, D. Saupe, T. Schreck, and D. Vranic, Automatic se-
lection and combination of descriptors for effective 3D similarity search,
in Proc. IEEE Int. Workshop Multimedia Content-Based Anal. Retrieval,
Dec. 2004, pp. 514521.
[145] A. Johnson, O. Carmichael, D. Huber, and M. Hebert, Toward a general
3-D matching engine: Multiple models, complex scenes, and efficient
data filtering, in Proc. 1998 Image Understanding Workshop (IUW),
Nov. 1998, pp. 10971107.

Yubin Yang (M04) received the B.Sc. degree in

computer science from the Wuhan Technical Univer-
sity of Surveying and Mapping, Wuhan, China, in
1997, and the M.Sc. and Ph.D. degrees in computer
Yao Zhang received the B.Sc. degree in English lan-
science from Nanjing University, Nanjing, China, in
guage and literature in 2004 from the Nanjing Uni-
2000 and 2003, respectively. versity of Technology, Nanjing, China, where she is
He is currently with the State Key Laboratory for
currently working toward the M.Sc. degree in the De-
Novel Software Technology, Nanjing University. He
partment of Information Management.
is also with the Joint Laboratory for GeoInformation
Her current research interests include information
Science, Chinese University of Hong Kong, Shatin,
retrieval, digital libraries, and digital publishing.
N.T., Hong Kong. His current research interests in-
clude multimedia information retrieval, particularly for images and 3-D models,
spatial data mining and knowledge discovery, medical image processing and
analysis, machine learning, and intelligent virtual geographical environment.