Classified 3d Model Retrieval Based on Cascaded Fusion of Local Descriptors

International Journal of Computer Graphics & Animation (IJCGA) Vol.6, No.1, January 2016
DOI : 10.5121/ijcga.2016.6102 10
CLASSIFIED 3D MODEL RETRIEVAL BASED ON
CASCADED FUSION OF LOCAL DESCRIPTORS
Konstantinos Mochament1
, Athanasios Tsakiris1
, Dimosthenis Ioannidis1
,
Dimitrios Tzovaras1
1
Information Technologies Institute Centre for Research & Technology Hellas, Thermi-
Thessaloniki, Greece
ABSTRACT
One of the core tasks in order to perform fast and accurate retrieval results in a content-based search and
retrieval 3D system is to determine an efficient and effective method for matching similarities between the
3D models. In this paper the “cascaded fusion of local descriptors” is proposed for efficient retrieval of
classified 3D models, based on a 2D coloured logo retrieval methodological approach, suitably modified
for the purpose of 3D search and retrieval tasks that are widely used in the augmented reality (AR) and
virtual reality (VR) fields. Initially, features from Key points are extracted using different state of the art
local descriptor algorithms and then they are joined to constitute the feature tuple for the respective key
point. Additionally, a feature vocabulary for each descriptor is created that maps those tuples to the
respective vocabularies using distance functions that applied among the newly created tuples of each Point
Cloud. Subsequently, an inverted index table is formed that maps the 3D models to each tuple respectively.
Therefore, for every query 3D model only the corresponding 3D models are retrieved as these were
previously mapped in the inverted index table. Finally, from the retrieved list by comparing the local
features frequency of appearance to the first vocabulary, the final re ranked list of the most similar 3D
models is produced.
KEYWORDS
Point Cloud, local descriptor, features tuple, keypoint, cascaded fusion.
1. INTRODUCTION
With the increasingly growing technology of augmented reality (AR) and virtual reality (VR) the
even greater growing technologies in the field of mobile devices, there is a major need of
computationally inexpensive algorithms that should be efficiently executed in limited resource
devices (mobiles ,tablets etc.) and seen as lightweight operators that perform AR and VR tasks.
One of the core tasks in the AR and VR field is the search and retrieval of similar 3D models
from a classified stored 3D repository. Therefore in this paper a lightweight matching method in
terms of processing, power and memory consumption is proposed that could be applied in both
limited resource devices or in cloud based infrastructures and could provide almost in real-time
the functionality of 3D model matching against a known 3D model dataset. Similar to the very
well-known internet text-based search engines where Keywords are matched with keywords from
different web pages and lead to web page retrieval, the local features describing Key points are
used in the same way. Therefore, Key point features are matched with Key point features from
different Point Clouds and lead to Point Cloud retrieval where each point Cloud is a
representation of a 3D model from a stored dataset. Providing a 3D model search and retrieval
capability to a low energy power and memory device could have unlimited possibilities and could
be applicable to several different scientific fields (i.e. robotics, healthcare etc.) that use 3D

International Journal of Computer Graphics & Animation (IJCGA) Vol.6, No.1, January 2016
11
modelling for simulation of the real word Use Cases and scenarios. Furthermore, due to the
decreasing cost of 3D sensors and in combination with the technological evolution of mobile
devices more and more commercial end user applications along with scientific research fields like
robotics and computer vision are observed to be built in such a manner that can support AR and
VR system functionalities. For that reason and in order to perform AR and VR operations, beyond
the accuracy of the performed tasks an important matter that should be always take into
consideration, as much as possible, is the reduction in complexity of the implemented algorithm.
Due to that in this paper we propose a method that takes into account the above limitations and
performs the majority of all heavy computation operations in a pre-process phase so that the task
of search and retrieval of 3D models will be performed in almost real time.
2. RELATED WORK
Numerous researches have been conducted and an equal number of methods have been proposed
for the task of searching and retrieving 3D models. More specifically, following state of the art
methods achieve high accuracy results evaluated against some of the most challenging datasets
like SHREC’13[1], ’14[2] and ’15[3]. The methods proposed include the hydric shape descriptors
[4] and meta similarity generation for non-rigid 3D model retrieval [6] that integrates geodesic
distance–oriented global features and curvature-oriented local descriptors and in a more general
approach can be applied to similar methods that integrates more than one features in order to
develop a specific algorithm for both partial and rigid 3D models. Another efficient descriptor
methods are the Histograms of Area Projection Transform, the Radical symmetry Detection and
the Shape Characterization with Multiscale Area projection transform [5] and theR-
BiHDM[2]method where it computes the project matrix k called reduced bi-harmonic distance
matrix and the resulted equation is equal to:
: lim
→∞

(,) (1)
Regarding the non-rigid a fast non-rigid retrieval method is the HKS-TS[7] where it constructs a
context-based integral kernel operator on a manifold and then applies modal analysis to map this
operator into low-frequency functional representation and lastly it computes the spectrum that
forms the shape based signature. Another efficient descriptor is the High level Feature Learning
for 3D Shapes, the Bag-of-Features approach with Augmented Point Feature Histograms, the BoF
and SI-HKS and the Spectral Geometry [8]. Regarding the retrieval algorithms evaluated in more
recent dataset the methods include the Cross-Domain Manifold for range Scan-based 3D Model
Retrieval, the 2D-DTC Coefficient of Silhouettes for Range Scan Matching, the Fast point feature
Histograms[3], the Bag of features and Modified Shape Distribution approach for Range Scan
Matching[3], the Partial Shape retrieval with Fast Point Feature Histograms(FPFH) and Signature
Quadratic Form Distance(SQFD) that is an adaptive similarity measure for content-based feature
representations of data, the Depth Image Similarity based method for range Scan Matching and
finally the Hough-voting in a Continuous Voting Space for Range Scan Matching[3].Regarding
the selected descriptors even though the proposed method could be applied in any pair of local
descriptors we select the PFH and the FPFH due to the fact that they were more suitable for the
developed engine. Therefore, the studied features [11] presented in the following table.

International Journal of Computer Graphics Animation (IJCGA) Vol.6, No.1, January 2016
12
Table 1. Descriptors important features
Descriptor
Number of
points
Number of
dimensions
Parallel Normals
PFH n 125 GPU Y
FPFH n 33 OMP+GPU Y
3. OVERVIEW OF THE PROPOSED METHOD
The proposed method framework is shown in Figure 1. The presented pipeline is based on an
image based method for logo retrieval over common pictures [13]. Based on the idea of the
referenced method for logo retrieval that uses cascaded sparse localized colour information we
suitably modify the necessary modules and with the addition of the following described
components the proposed method is formed for search and retrieval over 3D models. Initially the
local features are extracted from the entire dataset using both FPFH and PFH descriptor
algorithms. Hence we create two distinct matrices for every 3D model, where each matrix is equal
to the number of the Key points of the respective 3D model and the size of the respective
descriptor’s dimensions. Then, from the extracted features of the dataset we create two local
feature vocabularies using the well-known BoF method. Subsequently, using a distance metric
function like the Euclidean distance we replace each feature from the two matrices of every 3D
model with the closest id from the features of the respective vocabulary. Subsequently an index is
created where for every id of each vocabulary a mapping is performed between the 3D models
that had at least once replace their feature with the respective id. By uniting the two indexes we
build a unified inverted index where for every tuple of ids we map only the intersection of the 3D
models that were previously mapped in each index.
Thereby for every query 3D model first features are extracted and then replaced with the closest
id as previously described. Next, from the inverted index only the 3D models mapped in the
corresponding tuples are retrieved. The final retrieved list is re ranked using the frequency of
appearance of the ids from the first vocabulary and a cosine similarity between these frequencies.
Figure 1.Cascaded retrieval framework

13
Even though the presented method could be applied in any local descriptors combination, for the
purpose of this research only the point feature histograms [12] and the fast point feature
histogram descriptors [9] are conducted.
Point feature histogram:
The Point feature Histogram (PFH) descriptor approach is based on computation of both surface
normal and also curvature estimation (xyz 3D data). To compute the relative difference between
two pointsp and q, first a reference frame is built on p using three vectors (u, v, w) where:
i. vector u is equal to the surface normal at p.
ii. vector v is equal to = × ( − )/ (2)and
iii. vector w is equal to ! = × (3)
Hence using the (u, v, w) the distance between p(np) and q(nq) is estimated by computing the
angles (α,φ,θ) where:
i. α is equal to = #$
( _) (4)
ii. φ is equal to ( = #$
( ( − )⁄ ) (5)and
iii. θ is equal to * = ## (! ∙ _ , ∙ _ ) (6)
Computing angles (α, φ, θ) for all the pairs in the k-neighbourhood of p, will result into a 125 bin-
histogram for each point p.
Fast point feature histogram:
Fast point feature histogram (FPFH) is the resulted descriptor algorithm after reduction
complexity of the PFH from O(nk2) to O(nk). This is achieved by compute the angles (α,φ,θ) only
between p and its k-nearest neighbours. This will result into the SPFH (simplified point feature
histogram) where the final FPFH is produced by the sum of the SPFH and the weighted based on
the distance to p value of its k-neighbour.
FPFH/ = SPFH/ +
2
3
∑ SPFH
3
562 (i)/w5 (7)
This way, each one of the three angles is described through an 11-bin histogram, where the
concatenation resulted in a 33-bin histogram. The above ensures that FPFH is invariant to scale
orientation and position
3.1. Pre-process
3.1.1 Phase 1: feature extraction and creation of vocabularies
Based on the well-known FPFH descriptor [9] in this phase first every 3D model is converted into
Point Cloud and then local features are extracted from all Point Clouds that are stored in our
database. This results into an NxM matrix where N is the number of key points and M is the size
of the descriptor which for fast point feature histogram (FPFH) is equal to 33 [9]. Having the
above matrix for every 3D model in our dataset a vocabulary is generated of visual words. These
visual words are the centroids of a clustering obtained using k-means, to the entire set of the
extracted features. For the k-means clustering, our experimental results show that the best
optimization is achieved whenk equals to 20% of the total features. Repeating the above step
using a different local descriptor we create a second NxM matrix where again N is the same
number of key points and M is the size of the descriptor which is 125 bins for Point Feature

14
Histogram [12].Additionally as previously described a second vocabulary of visual words is
created again by obtaining k-means clustering to the newly created features dataset.It is worth
mentioning here that the selected descriptors could be replaced with any other local feature
descriptors
3.1.2 Phase 2: mapping of the 3D models against the vocabularies’ ids
Mapping the two matrices of each 3D model to the two vocabularies created in phase 1 the
matrices are encoded by replacing each one of the feature with the closest visual word of the
respective vocabulary. Hence the two matrices are replaced with the ids of the closest visual word
from the vocabulary. If N is the total number of the first descriptor set and K is the total number
of the respective visual word vocabulary, then for every ith (i=1,2,3,….N) we perform a brute
force matching against the K visual words and we select the jth(j=1,2,3….K) visual word id that
has the minimum distance d to the ith feature. As a distance metric we use the simple Euclidean L
distance where:
L = d(p, q) = (p2 − q2)= + (p= − q=)= + ⋯ + (p? − q?)= (8)
Therefore every ith feature is replaced with the id of the jth visual word.
3.1.3 Phase 3: Inverted indexing of the 3D models dataset
In this phase we extend the created vocabularies by mapping the 3D models that have at least one
appearance in the respective visual word of every vocabulary. Therefore having the resulted
matrices of the previous phase we link the 3D models of our dataset with respect to the ids of the
visual words. Finally, from the above we create an inverted index table consisted of visual word
id tuples and the intersection of 3D models between the two visual word ids. If for instance {id1,
id7, id567,…,idm} is the vi encoded descriptor of the ith 3D model the inverted index will map
the ith 3D model only to the visual words with the ids: ida1, ida7, ida567,…,idam respectively.
Doing that for all the 3D models of the dataset and then form the inverted index will result in a
table of ids where each id will be linked with the 3D models that has at least one appearance, for
example let idai is the following set {3Dmodel1,3dmodel5,…, 3DmodelN etc.}. Repeating the
above for the second (b) visual word vocabulary and joining the two inverted indexes will result
into an NxM table, where N is the size of the first vocabulary and M is the size of the second. To
form the final inverted index table we only need to include for every possible tuple of ids the
intersection of their linked 3D models. For instance, if{3Dmodel1, 3Dmodel5, 3Dmodel897}is the
idai set and{3Dmodel7, 3Dmodel11, 3Dmodel5}is the idbi set. Then, the intersection of these two
sets is described by the following equation:
(Idai ∩ Idbi) = {E: E ∈ # ˄ E ∈ H} (9)
Hence, in the final inverted index table only the 3Dmodel5 will be linked to that tuple because it’s
the only 3D model left after the performed intersection of the two sets. Therefore, equation (9)
will be equal to:
(Idai ∩ Idbi) = {3Dmodel5} (10)
3.2 Phase 4: Query 3D model processing
Having a query model its set of features is extracted using the two different descriptors thatwere
also used in the first phase for the entire dataset and then the two matrices are encoded by
replacing the features with the closest visual word id from the respective vocabularies. Hence by

15
the union of these two matrixes we constitute a new query table that consists of visual word ids
tuples.By comparing this query table to the inverted index table we retrieve only the 3D models
that appear in the respective tuples eliminating the rest of the dataset.For instance if the closest
tuple of ids for a single keypoint from the query 3D model is: (idan, idbm) then from the inverted
index table we find the corresponding tuple and retrieve the linked 3D models.
3.3. Phase 5: Re ranking of the retrieved list
The final list of 3D models is re-ranked using as a distance metric the cosine similarity function
betweenthe frequencies of appearance histograms based on the visual words vocabulary created
from the first descriptor.
s(A, B) =
∑ RSTS
U
SVW
X∑ RS
Y
U
SVW X∑ TS
Y
U
SVW
(11)
Adding the weighted class distance
As an additional final step and using the information of the predefined categorization of the
database 3D models we can achieve significantly better optimization in the task of category
recognition by taking into account the distance between the query model and each and every class
of the dataset. Adopting the proposed method [15] we compute this step and weight it accordingly
so that the final list generated based on the distance described above and the class distance
computed in this phase. As proposed, to measure the class distance we can use the centroid the
average or the minimum distance. Since average and centroid methods need more computations
we use the minimum distance class.
dZ[q, C]^ = min?∈`a
{d(q, m)} (12)
The selection of the minimum against the other two methods is based on the fact that since it is
the distance between the query 3D model and the closest 3D model from the respective class it
has already been computed in the previous phases. Hence without adding any complexity to our
method we integrate the class distance information in our final results. Lastly, for the
computations of the final distance in order to identify the query 3D model category based on the
proposed method [16] we adopt the following equation.
d[q, C]^ = [c
2
3
d ∑ dZmin(
562,=,….,3 q, C])] (13)
Where Cj is the jth class, k is the kth 3D model of the ranked list and dc is the respective class
distance. The weights (k=1/k) are linearly decreasing having as a starting point the first 3D model
in order to give more weight in the class of the closest 3D models.

Figure 2. Cascaded retrieval framework
4. EXPERIMENTAL SETUP AN
Database and setup
In order to evaluate the proposed metho
initial set was created by CERTH/ITI
Informatics, Telematics and Telecommunications, with long experience in
and national RD projects. Additionally, usin
was created with higher quality 3D models in order to evaluate our method in a second more
challenging dataset. The first dataset consist of
categories and the second training set consist of 11
use two sets, approximately equal to
models respectively. The two complete sets are
Momiras project [14]. Regarding the down sampling process, all features were extracted after
sampling the Point Clouds with a grid size of 1.0cm.
Figure 3.Retrieval results
Figure 2. Cascaded retrieval framework
XPERIMENTAL SETUP AND BASELINE
proposed method, a subset from a large 3D models dataset was used. The
created by CERTH/ITI [16] one of the leading Institutions of Greece in the fields of
Informatics, Telematics and Telecommunications, with long experience in numerous European
Additionally, using the Structure Sensor [15]a second smaller dataset
with higher quality 3D models in order to evaluate our method in a second more
challenging dataset. The first dataset consist of a training set of 565 3D models from 10 different
econd training set consist of 110 high quality 3D models. For the test sets we
equal to the one third of the training sets, namely 100 and 45 3D
ively. The two complete sets are publicly available through the web
. Regarding the down sampling process, all features were extracted after
sampling the Point Clouds with a grid size of 1.0cm.
Retrieval results from a search performed using a chair 3D model as a query
16
d, a subset from a large 3D models dataset was used. The
one of the leading Institutions of Greece in the fields of
numerous European
a second smaller dataset
with higher quality 3D models in order to evaluate our method in a second more
a training set of 565 3D models from 10 different
0 high quality 3D models. For the test sets we
the one third of the training sets, namely 100 and 45 3D
through the website of the
. Regarding the down sampling process, all features were extracted after
a query

17
Baseline and metrics
In order to evaluate our proposed method we compare our results with the retrieval results,
achieved by each of the descriptors algorithm separately. Hence since our proposed method uses
the combination of FPFH and PFH descriptor methods the performance of the two descriptors
separately was measured as well. Moreover, the same pipeline used for the proposed method was
adopted again with the only difference that after the extracted descriptors direct similarity
operations were performed without the further pre-process steps. Finally to define a similarity
metric between two sets of extracted descriptors the following distance metric was used where
that is the sum of the distance between the centroids and the distance of their standard deviation
respectively [11].
D(A, B) = L2(cR, cT) + L2(stdR, stdT) (14)
stdR(i) = X 2
|R|
∑ ca](i) − cR(i)d
=
|R|
]62 , i = 1,2,3, . . n (15)
L2(a, b) = ∑ |a(i) − b(i)|
m
562 (16)
Evaluation metrics
To evaluate our proposed method we compare the methods using the Mean Average Precision
where the Average precision achieved by each query is computed for the total number of the
performed queries

no =
∑ pqrs(t)
u
vVW
w
(17)
5. RESULTS
Table 12 presents the experiments’ results where the retrieval methods were ranked according
to the mAP score they achieve.
Table 1. : First set mean average precision (MAP) scores.
Method Map score
PFH 0.386
FPFH 0.357
Cascaded local descriptor 0.486
Table2. : Second set mean average precision (MAP) scores.
Method Map score
PFH 0.537
FPFH 0.515
Cascaded local descriptor 0.607
Table 34 presents the category recognition based on the overall accuracy of the performed
queries
Table 3. : Category recognition accuracy of the first set.
Method Category Recognition
PFH 39%
FPFH 36%
Cascaded local descriptor 49%

18
Table 4. : Category recognition accuracy of the second set.
Method Category Recognition
PFH 54%
FPFH 52%
Cascaded local descriptor 61%
6. CONCLUSION
Having a 2D-based methodological approach for coloured logo retrieval over image datasets as a
reference point, this retrieval approach is expanded in the 3D field by applying the suitable
transformations and the appropriate additions in order to develop the proposed 3D search and
retrieval method. According to this, a shape based retrieval approach for 3D models was
presented. The proposed method takes advantage of more than one local descriptor algorithms
and by creating several lists of 3D models for every tuple of local features, it decreases the final
number of retrieved 3D models. Hence by performing similarity metrics only to the final
significantly smaller list of 3D models we achieve fast and accurate results. Additionally the pre-
processed inverted index ensures fast retrieval without a one on one (linear complex) comparison
on all of the stored 3D models but instead a selective retrieval based only on the corresponding
subsets. Furthermore, the usage of the classification information of the initial models in the two
newly created 3D model datasets that are publicly available under the Momiras project official
page ensures a significant optimization in the task of object category recognition. Finally, in
future workthe presented methodology will be applied using different local feature extraction
algorithms and using also different 3D model datasets in order to confirm and refine the presented
approach.
Acknowledgements
The research leading to these results has received funding from the Momiras bilateral Greece-
Israel 2013-2015 research and technology collaboration project.
Figure 4. Precision Recall on the first dataset
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1
Fpfh
pfh
cascaded

19
Figure 4. Precision Recall on the second dataset
REFERENCES
[1] LI, Bo, et al. SHREC'13 track: large scale sketch-based 3D shape retrieval. In: Proceedings of the
Sixth Eurographics Workshop on 3D Object Retrieval. Eurographics Association, 2013. p. 89-96.
[2] LI, Bo, et al. Shrec’14 track: Large scale comprehensive 3D shape retrieval. In: Eurographics
Workshop on 3D Object Retrieval. 2014. p. 131-40.
[3] GODIL, A., et al. SHREC’15: Range Scans based 3D Shape Retrieval.
[4] LI, Bo; GODIL, Afzal; JOHAN, Henry. Hybrid shape descriptor and meta similarity generation for
non-rigid and partial 3D model retrieval. Multimedia tools and applications, 2014, 72.2: 1531-1560.
[5] GIACHETTI, Andrea; LOVATO, Christian. Radial symmetry detection and shape characterization
with the multiscale area projection transform. In: Computer Graphics Forum. Blackwell Publishing
Ltd, 2012. p. 1669-1678.
[6] YE, Jianbo; YAN, Zhicheng; YU, Yizhou. Fast nonrigid 3D retrieval using modal space transform.
In: Proceedings of the 3rd ACM conference on International conference on multimedia retrieval.
ACM, 2013. p. 121-126.
[7] SUN, Jian; OVSJANIKOV, Maks; GUIBAS, Leonidas. A Concise and Provably Informative
Multi‐Scale Signature Based on Heat Diffusion. In: Computer graphics forum. Blackwell Publishing
Ltd, 2009. p. 1383-1392.
[8] LI, Chunyuan. Spectral Geometric Methods for Deformable 3D Shape Retrieval. 2013. PhD Thesis.
Concordia University.
[9] RUSU, Radu Bogdan; BLODOW, Nico; BEETZ, Michael. Fast point feature histograms (FPFH) for
3D registration. In: Robotics and Automation, 2009. ICRA'09. IEEE International Conference on.
IEEE, 2009. p. 3212-3217.
[10] OHBUCHI, Ryutarou; FURUYA, Takahiko. Accelerating bag-of-features sift algorithm for 3d model
retrieval. In: Proc. SAMT 2008 Workshop on Semantic 3D Media (S-3D). 2008. p. 23-30.
[11] ALEXANDRE, Luís A. 3D descriptors for object and category recognition: a comparative evaluation.
In: Workshop on Color-Depth Camera Fusion in Robotics at the IEEE/RSJ International Conference
on Intelligent Robots and Systems (IROS), Vilamoura, Portugal. 2012.
[12] RUSU, Radu Bogdan, et al. Aligning point cloud views using persistent feature histograms. In:
Intelligent Robots and Systems, 2008. IROS 2008. IEEE/RSJ International Conference on. IEEE,
2008. p. 3384-3391.
[13] PANDEY, Rohit, et al. Cascaded sparse color-localized matching for logo retrieval. In: Image
Processing (ICIP), 2014 IEEE International Conference on. IEEE, 2014. p. 2207-2211.
[14] 160.40.50.50/Momiras,. Momiras Content Import Toolkit. N.p., 2015. Web. 3Nov. 2015.
[15] Structure.io,. Structure Sensor - 3D Scanning, Augmented Reality, And More For Mobile Devices.
N.p., 2015. Web. 3Nov. 2015.
[16] Iti.gr,. ΙνστιτούτοΤεχνολογιώνΠληροφορικήςΚαιΕπικοινωνιών. N.p., 2015. 3 Nov 2015.
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1
Fpfh
pfh
cascade
d

20
Authors
Short Biography
Mochament Konstantinos is a postgraduate student in Information and
Communication Technology (ICT) Systems of International Hellenic University (IHU)
and from November 2014 is working as a Research Associate in Informatics and
Telematics Institute of the Centre for Research and Technology Hellas in Thessaloniki
(CERTH/ITI). His main research interests are in the areas of 3D data processing.
Thanos Tsakiris is a Research Associate in the Informatics and Telematics Institute.
He received the Diploma in Computer Science from the Aristotle University of
Thessaloniki, Greece and the MSc in Computer Games Technology from the
University of Abertay Dundee, UK in 2000 and 2001 respectively.His main research
activities include 3D rendering, VR/AR/MR technologies and applications, Multimedia
Authoring and Multi-modal Human-Computer Interfaces. His involvement in these
research areas has led to his participation in the authoring of 2 articles in refereed
journals, 2 papers in international conferences and the co-authoring of 1 book chapter.
Mr. Dimosthenis Ioannidis is working from October 2006 as a Research Associate in
Informatics and Telematics Institute of the Centre for Research and Technology Hellas
in Thessaloniki.His main research interests are in the areas of biometrics (gait and
activity-related recognition), 3-D data processing, and web semantics.
Dimitrios Tzovaras is a senior researcher (Grade A’) and the directorof the Information
Technologies Institute (ITI) and a visiting professorof the Imperial College London. His
main research interests includevisual analytics, computer vision,knowledge
management, multimodalinterfaces, and virtual reality. His involvement with those
researchareas has led to the co-authoring of over 80 articles in refereed journalsand 200
papers in international conferences. He has served as a regular reviewer for a number of
international journals andconferences.

Classified 3d Model Retrieval Based on Cascaded Fusion of Local Descriptors

More Related Content

What's hot (20)

Similar to Classified 3d Model Retrieval Based on Cascaded Fusion of Local Descriptors (20)

Recently uploaded (20)

Classified 3d Model Retrieval Based on Cascaded Fusion of Local Descriptors