Location-dependent Content-based Image Retrieval System
Based on a P2P Mobile Agent Framework
Yusuke Ariyoshi Faculty of Economics, Management and Information
Science, Onomichi City University
Onomichi, JAPAN email: [email protected]
Junzo Kamahara, Naoki Tanaka, Katsutoshi Hirayama, Takashi Nagamatsu
Graduate School of Maritime Sciences Kobe University Kobe, JAPAN
email: {kamahara, ntanaka, hirayama}@maritime.kobe-u.ac.jp
Yuuichi Teranishi Cyber Media Center,
Osaka University Suita, JAPAN
email: [email protected]
Abstract—In this article, we propose a geoconscious content-based image retrieval system based on a P2P mobile agent framework. This system retrieves similar photographs from an image database of location-dependent photographs (e.g., photos of buildings, landmarks, etc.), which use GPS positions for geo-tagging. The P2P mobile agent framework supports intelligent agents. This agent searches similar content image using a query photograph by traversing the P2P network, instead of the mobile device issuing the query. In this paper, we describe the design of the proposed system and a portion of its implementation. This prototype system produces a new peer and rearranges the placement of image agents among peers for workload balancing. Furthermore, we provide the experimental results of our implementation for managing location-dependent image agents, which are clustered with peers in a distributed Delaunay network.
Keywords-component; P2P; Content-Based Image Retrieval; Geolocation; Geo-tagging; Geoconscious; Delaunay network
I. INTRODUCTION
Based on the success and growth of photo-sharing sites such as Flickr, a large number of photographs are being shared over the Internet. The number of uploaded images is increasing rapidly, and an even greater number of images will be shared in the future owing to the increasingly widespread use of mobile devices that have cameras, such as smartphones. Searching similar photographs from such a huge image database is a typical problem in content-based image retrieval (CBIR).
These photographs can be classified into a number of types, including portraits, landscapes, artistic, and documentary photographs. The target of our research is retrieving location-dependent photographs (e.g., photographs of buildings or landmarks). The location of a photograph is indicated by the GPS position, i.e., the latitude and longitude
where the photographer has taken the picture. This embedding of locational information into a photograph is known as geo-tagging.
We studied a method for retrieving similar images efficiently by utilizing the geolocation of images. It is efficient to limit the geographic range when searching for a location-dependent image. Using a combination of the geographic distance and image distance (using image features) also increases the performance of CBIR. We call this type of process “geoconscious” CBIR [1]. In this paper, the term geoconscious means that the geolocation is the primary criterion used for the retrieval of images.
For image retrieval, calculating various image features and comparing their high-dimensional vectors requires a significantly high calculation cost. While the calculation cost of image pairing is relatively low, comparing a photograph with a vast image dataset to find a similar image requires a large amount of computation. Therefore, as such workload increases, the efficiency of our retrieval method has to increase in-scale with the computation power required.
Such a complex CBIR requires intelligent behavior for each entity used in calculating the similarities. The main concept underlying geoconsciousness is the management of intelligent photo agents based on their locations on the P2P network. As intelligent agents on the P2P network work in a scalable manner, finding similar images using image features and geolocations is a suitable method for a mobile agent-based P2P framework. Structured P2P networks exhibit good usage efficiency for a specific search range. Therefore, we adopted P2P Interactive Agent eXtension (PIAX) [2] as such a framework.
PIAX is a structured P2P framework with an enhanced location-oriented service. It can handle intelligent agents that move across a structured overlay network with multiple peers. Agents can be implemented to calculate various types
978-1-4244-9529-0/13/$31.00 ©2013 IEEE
The 9th International Workshop on Mobile Peer-to-Peer Computing 2013, San Diego (18 March 2013)The 9th International Workshop on Mobile Peer-to-Peer Computing 2013, San Diego (18 March 2013)
72
os
A
sgrg
iAlnGptlpurictup
nIsoprr
aioapqsaiagaties
of features. Fusearch for the
A. Related W
For P2P surveys [3-5] goal of theseretrieving imageolocation.
For imageimages has emA survey in [landmark reconot directly reGPS position position from the content olandmark but people, food, auses keypointrecognition neimage and noclassified landthe use of geusing the Gpreviously kno
The remainnext section ouIn section 3, system. In secof our systempreliminary exremarks regarresearch in sec
II. GEO
In this secarchitecture. Ois a P2P netwoverlay netwoagent creationprovides a locquery. The useveral types application sincludes an imand its meta dgeographical agent (CrA), wthe Internet; interface of oexecutes a simshows a diagra
urthermore, PIagents on its n
Works
multimedia son CBIR fro
e studies wasages or multim
e retrieval, thmerged as a re6] introduced
ognition. The gepresent the lanis not that of where the phoof a photogrmay also incluand souvenirst (local featureeds local featot global featudmarks in largeolocations, a
GPS positionsown landmarknder of this arutlines the conwe describe
ction 4, we desm, and in secti
xperiment. Firding this studction 6.
OCONSCIOUS PBASED IMA
ction, we giveOur system utiwork layer, w
ork. The peersn, to the uppcation-dependeupper layer is
of mobile ageservices. Our
mage agent (IAdata; a cluster region and IAwhich crawls a web agent
our system; amilar image retam of the relat
IAX supports network.
searches, thereom P2P netwos to create gmedia data wit
he geo-tagginsearch field ovthe concepts
geolocation ofndmark in thethe landmark
otograph was craph might nude various ot. Therefore, lare) matching, ture matching ure matching
ge-scale image and [8] constrs and camerks. rticle is organincept of geocothe suitable
scribe some ofon 5, we discinally, we givdy and descri
P2P MOBILE AAGE RETRIEVA
e an overview ilizes two layewhich consists provide agenper layer. Thent query funcs the agent laents work togr geoconscio
A), which managent (ClA),
As within the to collect geo(WA), whic
and a query atrieval query ftionships amon
a geometrical
e have been orks. Howeve
generic systemthout regard to
ng of photogrver the past deof geolocatio
f a photographe image, becau
itself but rathcaptured. Morenot include other objects, suandmark recog
because landof the object of the image
collections wructed a 3D mra parameter
ized as followsonscious P2P Cagents used if the noted behcuss the resultve some conclibe potential
AGENT CONTEN
AL
of our P2P sers. The lowerts of peers annt functions, suhe overlay nection, such as ayer. In this gether to realizous CBIR snages the imag
which managregion; a cra
o-tagged imagch is the webagent (QA), wfrom the user. ng these agent
range
short er, the ms for o their
raphic ecade.
on and h does use the her the eover,
only a uch as
gnition dmark in the
es. [7] without
model rs for
s. The CBIR. in our havior ts of a luding future
NT-
system r layer nd an uch as etwork
range layer,
ze the system ge date ges the awling ges on b user which Fig. 1 ts.
A.
Ton useothimagen
B.
TandpreNatthe
T1) C
2) Fth
Ip
3) C
bfa
4) S
fT
discby imaamoQAretr
C.
Tpeeguaregexa
Figure 1. Rela
Crawling Age
The CrA crawthe Internet. T
es Flickr, Paner sites. The Cage and othernerated IA will
Image Agent
The IA has fund managing thsent. Each IA turally, the numnumber of im
The functions oCalculate the im
The IA shouFind and movehe geolocation
Each ClA mIA asks the PIAposition of the Calculate the s
The IA is rebetween the quformula that spand geolocatioStore and recov
PIAX has a features to andThe IA shouldcovery of agenother agents.
ages should bount of netwo
A moves to a prieval.
Cluster Agent
The ClA manaer as a Voronoaranteed to beion than to an
ample of a Vor
ationships betwee
III. AGEN
ent
wls to collect shTo obtain the pnoramio (httpCrA then prodr textual or l move to the p
nctions for calche geo-tag ancorresponds tmber of IAs in
mages that can bof the IA are amage features uld contain lowe to the peer thn of the image manages a diffAX for the proimage. imilarities bas
equested to calueried image apecifies how t
on into a similaver the featurepersistent fun
d from the nodd only includents and the ran
In our architbe located in ork traffic for peer containing
t
ages the geomoi region. In ae closer to theny other kernronoi diagram
en different agents
NT DESIGNS
hared geo-tagphotographic
p://www.panorduces a new IAgeo-tagged inproper peer wi
culating severand textual ino one image on a P2P netwobe searched.
as follows:
w-level featurehat manages th
fferent geograpoper ClA usin
sed on the querlculate the degand the managto assemble tharity score. es to/from stornction to storee storage.
e simple functinking of imagetecture, IAs ththe same peequery messag
g the candidat
metric region oa Voronoi rege kernel point nel point [9]. F.
s in our system
ged photograpimages, the C
ramio.com), aA to manage tnformation. Tithin its locatio
al image featunformation whof the photograork is the same
es of the imagehe locations n
phic region. Tng the geograph
ried featuresgree of similared image usin
he image featu
age e and recover t
ions because tes are performhat have simier to reduce tging because tte IAs before t
of the IAs in tion, any pointof that Voron
Fig. 2 shows
phs CrA and the
The on.
ures hen aph. e as
e. ear
The hic
rity ng a ures
the
the med ilar the the the
the t is noi an
The 9th International Workshop on Mobile Peer-to-Peer Computing 2013, San Diego (18 March 2013)
73
tdtp
bnCmdap
D
WWu
pma
noeiq
E
u
F
PIAX supptriangulation diagram. The the locations points from th
Furthermorebalance of thenumber of IAClA will splitmanages half described in approach this problem [10].
D. Web Agen
The WA isWA can existWA and inputupload photog
The WA wipage when thmay be impleasynchronousl
As there is not limited toonly the P2Pensures the scin our architecquery each IA
Fi
E. Query Age
The QA exeuser. The func
1) CalculateThe QA
Figure 2. Examp
ports a Delaudescribes theClA calculatesof its own k
he peer. e, the ClAs peers. When
As in the peer t into two ClA
f of the existia later sectioproblem as a
t
a web user it on multiple t a photographgraphs, a QA iill display the
hey are returnemented usingly. no central m
o the use of a P network of calability of thcture, because
A of different p
igure 3. Scalabil
ent
ecutes a similactions of the Qe image featureA also has to c
ple of a Voronoi d
unay overlay e dual structs the Voronoi
kernel point an
should optima new IA is adis above a ce
As and produing IAs. The on. In future a generalized
interface on apeers, and us
h through a wes created. ranked image
ed from the Qg Ajax to show
management ofsingle WA. Tour system
e image searche the QAs prodpeers in paralle
lity of multiple us
ar image retrieQA are as follow
es contain feature
diagram
network. Delture of a Voregion by acqnd neighbor k
mize the wordded to a peer,ertain thresholuce a new pee
details of thresearch, wemutual assign
a P2P networksers can acceeb page. After
e results on thQA. The webw the image r
f the agents, wThe WAs mus(See Fig. 3).h and retrievaduced by a Wel.
ser search
eval query frows:
es of the image
launay oronoi
quiring kernel
rkload , if the ld, the er that his are e will nment
k. The ss the r users
he web b page results
we are st join . This
al used WA can
om the
e.
2
3
4
5
appfunof resu.
A.
InstruimaFuralloordgeoClAwor
Tnewan tocalcto imageothenwor
2) Find and mnear the geo
This funct3) Form the ran
The QAgeolocatioretrieved
4) Obtain the sThe QA
IAs to calcuof the QA an
5) Rank the IATo return
QA must ordA QA is a s
pear very similnctions to obtaimages. The ults to the WA
IV. I
Constructing
nitially, we uctured P2Page features rthermore, cluows the imageder. Each clustolocation. TheAs also assign rkload.
The CrA collew IAs to manaIA is created,determine t
culating the imfind a properage. One Clographically con moves into trkflow of this
Figure 4.
move to the pelocation of thetion is the samnge query for aA forms and son of the imagcandidates.
similarity scoreissues the qu
ulate the similand IAs.
As based on thethe similarity
der the imagessubclass of anlar to those of ain the similar
QA will be A.
IMPLEMENTIN
the Image Da
need to formnetwork. Thi
extracted bustering the imes to be collectter is represene ClAs manag
the IAs to a s
ects geo-taggeage the images the IA shoul
the similaritiemage features,r peer accordilA exists onorresponding Cthat peer. PIAXis shown in Fi
Workflow of im
eer that manage image
me as that in thagent retrievalsends a range ge, and receive
e from each reery message arity score bet
eir similarity sy score resultss based on thein IA. The funan IA. Howev
rities and calcudismissed aft
NG SYSTEM BEH
atabase
m an image is database cby analyzing
mages based oted based on tnted as a ClA ge the geomespecific peer f
ed photographs and other infd calculate thes to other , the IA will ring to the ge
n each peer. ClA will answX supports moig. 4.
mage database con
ges the locatio
he IA. l query using t
es the IAs for t
etrieved IA to the candid
tween the imag
cores s to the WA, tir score. nctions of a Qver, a QA has tulate the rankiter returning t
HAVIOR
database oncontains seveg the imagon their geo-tatheir neighbori
according to etric region. Tfor balancing t
hs and producformation. Whe image featu
images. Afrequest the Clolocation of t
Therefore, twer the IA, whobile agents. T
nstruction
ons
the the
date ges
the
QA the ing the
n a eral ges. ags ing its
The the
ces hen ures fter lAs the the ich
The
The 9th International Workshop on Mobile Peer-to-Peer Computing 2013, San Diego (18 March 2013)
74
B
tocp
ptTToprkwnhrrrn
B. Splitting re
When a nethe number ofof IAs, whichcertain threshproduce a new
Figure 5.
When a nepeer, a methodthe gravity cenThe gravity Therefore, weof the existingpeer will be crregion will thkernel points. will be creatednot be changehas the new krelationship recalculating trearrangementneighbor peers
Figure
Pe
Image A
Region Split
egion for load
ew IA is to bef IAs already eh indicates thehold, the ClAw peer for load
Balancing the wo
ew peer is prod for dividing nter of the IAcenter may
e can obtain a ng kernel point treated with thhen be divideThe new boun
d. The region ed without thikernel point wwith the o
the Voronoi ret of the imas (Fig. 6).
6. Relation betw
eer
Cluster Agent
Agents
Age
Commuadjustin
Productiowhen ab
d-balancing
e added to a peexisting in the e capacity of
A will split ind balancing (Fi
orkload of peers b
oduced by splithe region mu
s in the regiondiffer from
new kernel pothrough the gr
his new kernel ed mid-perpenndary of the twthat has the os new bounda
will be changedther kernel
egion of the neage agents th
ween new and old
ClusteAgen
I
ent Move
unication for ng workload
on of new peer bove threshold
eer, the ClA cpeer. If the nuthe peer, is o
nto two ClAig. 5).
by moving the IA
itting an overlust be chosen.n will be calcu
the kernel oint as the reflravity center. Apoint. The Vo
ndicular of thwo Voronoi reld kernel poin
ary. The regiod based on th
point. Therew peer will cahat belong i
d kernel points
Peer
er nt
Image Agents
checks umber over a
As and
As
loaded First,
ulated. point. ection A new oronoi e two egions nt may on that e new refore, ause a n the
C.
netwFignetwpee
1) LploOfr
2) Dnthth
3) Nns
D.
inpufromfor PIAranper
imacalcagesevissuusinTheandrela
123456789
1
1
Distributed D
A Delaunay work layer by
g. 7 shows a work is made
er:
Figure 7. Algo
Local Delaunproper Delaunocation inform
Other peers arefrom the calculDelegation: Anon-neighborshe nearest nehen deletes inf
Notification of neighbor is toltep 1, to assur
Retrieve Simi
When a userut a query imam the query imPIAX in the f
AX will answenge. The value rformance, whi
When the QAages managedculating the sient is performveral methods ues a query wng image featue QA then sord presents theationship betw
1. List NodeL2. List Neigh3. List NonN4. while(true)5. clearAll 6. addAll N7. clearAll8. sortAsCl9. localDela
Neighb10. delegatio
NonNei11. notificati
Delaunay Netw
overlay netwy an autonomosummary of t
e by repeating
orithm of distribut
ay Triangulanay network mation of the oe divided intolation results.
A peer does n. Information eighbor from formation of tf Triangulationld of its adjacre communica
ilar Image.
r searches for age into the Wmage. The QAform of a ranger with the age
of the query rich we will disA receives thed by the ageimilarities betw
med for each for calculating
with a formulures and the gerts the similare results to
ween agents is s
List hborNodeList NeighborNodeL
NonNeighborNeighborNode NeighborNodlockwise NodaunayTrianguorNodeList, N
on(NeighborNighborNodeLiionOfTriangul
work Generatio
work is generaous distributivethis algorithmg the followin
ted Delaunay netw
ation: A peeraround itself
other peers thao neighbors an
not require inof a non-neighthe non-neig
the non-neighbn: During this
cent peers, as ration between t
a photographWA. The WA thA constructs thge query basedents that exist range radius ascuss later. e list of agentnts), the procween the queragent on the
g an image simla for calculateographic distrity informatiothe user. A shown in Fig.
List
rNodeList eList to NodedeList deList ulation(NodeLiNonNeighborNNodeList, ist) lation (Neighb
on
ated at the Pe algorithm [1
m. The Delaunng steps on ea
work generation
r calculates tf based on tat this peer hond non-neighbo
formation ofhbor is passedghbor. The pbors. s procedure, earecognized frothem.
h, he/she shouhen creates a Q
he query messad on the locatiowithin the que
affects the sear
s (i.e., candidcedure used ry image and te list. There milarity. The Qting a similartance into the Ion from each diagram of t8.
eList
ist, NodeList)
borNodeList)
P2P 11]. nay ach
the the lds. ors
its d to eer
ach om
uld QA age on. ery rch
date for the are QA rity IA. IA the
The 9th International Workshop on Mobile Peer-to-Peer Computing 2013, San Diego (18 March 2013)
75
E
iifgfd“iapoosT
bs
sdc
A
uwGsmntigt
Figur
E. Ranking F
Considerinimage improvimage retrievafunction, whigeographic disfunction that udistance. The “orthologous”image as a simas a building opropose the uorthologous orthologous idsimilarity betwThe definition
),( cq IPPOIF
),( cq PPFOI
GEODist(P
GE
where d(Pbetween Pq anspecified by th
We develosearch systemdatabase conscalculations, a
A. Image Dat
For an imaused the 64-bwith an Intel GB of memorsearches for amoves to the pnumber of IAthe IAs re-alloimage featurgeolocation dathe PIAX fram
re 8. Workflow
Function
ng the distanceves the performal. Our researich combinesstance, has a buses only the goal of our i
” to the query milar image thor landmark, cuse of a conjuidentity func
dentity distancween the queryns of OIFs are
,( cq PPIMGDist
(5.0 PIMGDist
(Pq, Pc) is defin
PPEODist q ,(
Pq, Pc) is thnd Pc, and R he user.
V. PRELIM
oped a mobilem. Thus far, wstruction funcand a comparis
tabase Constr
age database cbit edition of CPU i7-3667Ury. The IA cra proper peer peer. The num
As is over the tocated. In thise calculationata were used mework. For th
for retrieving a si
e between quemance of the rch [12] shows the image better performimage distancimage retrievaimage. We de
hat captures thcontained in thunctive rankinction (OIF), ce, which reprey image and thshown below:
exp() c GEOD
5.0), cq GEPP
ned as follows
R
PPdP q
c
,()
he physical grepresents the
MINARY EXPER
e agent P2P gwe have impltion using PIson function.
ruction
construction teWindows 7 o
U 2.0 GHz 2Created from thaccording to
mber of IAs is threshold, the s test, the IA
n or a compto check the phis, we used 1
imilar image
ery image and location-depe
ws that our radistance wit
mance than a race or the geogral is to find imefine an ortholhe same objecthe query imagng function, i.
to calculateesents the deghe candidate i:
1)),( cq PPDist
),( cq PPEODist
s: Pc )
,
geographic dise radius of ret
IMENT
geoconscious lemented an IAX, image f
est using PIAXon VMware PCore/4Thread he geolocationits geolocatiothen checked. peer is dividedid not proceparison, andperformance a100,000 geolo
target endent anking th the anking raphic mages logous t, such ge. We .e., an e the
gree of mage.
(1)
(2)
(3)
stance trieval
image image
feature
X, we Player and 4
n data on and
If the ed and ess the
only around cation
data9. Tranpee
100was226peethatdataDelgray
B.
Pro3.3 proIAsandThifivedom
ThemayThe
C.
mayno as sTheof Othranthe unrthisima
comdist(1) shothe
asets from FlicThe location o
ndom number ter for load bala
For an avera0,000 geolocas 441.7, and 6.6. When a pers from the net was dividedabase construlaunay overlay
ay polygon ind
Figure 9. V
Image Featur
For this test, ofessional on V
GHz 2Core/ogram read thes on a peer. Thd calculate theis equates to e image featurminant color, eIt took 5.12 s
erefore, one IAy have only erefore, we mu
Search Radiu
A location-dey be retrieved errors in the csmall as possierefore, the se
the locationherwise, a smange. On the oth
performancerelated imagess way, there iage retrieval pFig. 10 sho
mparing the ptance and colo
while OIF(2own as the mea
image distanc
ckr. The initiaof the initial pto every trial. ancing was 500age of ten triaations. After th
the average npeer was divieighbor peer,
d. Fig. 9 showction algorithmy network, wh
dicates a Voron
Voronoi regions an
re Calculation
we used the VMware Playe/4Thread and e URL list ofhe sequential e MPEG-7 im0.271 s per imres for the testedge histogramto compare a
A processed t18,000 IAs wust test multipl
us vs. Retrieva
ependent imad based on its ocaptured GPS ible. Howeverearch radius nn within theall radius losesher hand, a lare of the imas will appear fis a tradeoff berformance.ows the resuerformance ofor histogram 2) uses (2). an reciprocal rce is a constan
al number of peers was chanThe threshold0 AIs.
als, it took 1,0he trial, the nnumber of IAided, 87.9 IAswith the excep
ws the results m. The blue lhich connects noi region man
nd Delaunay over
n and Compari
64-bit editioner with an Inte
8 GB of mef Flickr and ptime taken to
mage features mage. We uset: color layout
m, and scalablequery with the
the query in 0within a 1 s ple peers for qu
al Performance
ge in a certaown GPS posidata, the sear
r, a GPS errorneeds to covere estimated s correct imagrge search radage retrieval from a large s
between the se
ults of our f OIFs agains(image distanThe performarank (MRR) [1nt value on th
peers was 3 × 3nged by addingd used to divid
062 s to procnumber of peAs per peer ws moved to nption of the pfrom our imalines represenpeers, and ea
naged by a ClA
rlay network
ison
n of Windowsel CPU i7-396emory. The tproduced 92,9read the imagwas about 7
ed the followit, color structue color. e 92,948 imag
0.055 ms. A pprocessing tim
uery processing
e
ain search radition. If there rch radius can r is unavoidabr a certain ext
GPS accuraes that are outius will decreabecause sevesearch radius. earch radius a
experiment t the geograph
nce). OIF(1) uance results 13]. The MRRhe radius becau
3 = g a
de a
ess eers was new eer age nt a ach A.
s 7 0X test 948 ges h. ing
ure,
ges. eer me. g.
dius are be
ble. ent
acy. t of ase eral In and
for hic ses are
R of use
The 9th International Workshop on Mobile Peer-to-Peer Computing 2013, San Diego (18 March 2013)
76
the image distance is unrelated with the geographic distance. Obviously, the performance of the OIFs overcomes both the geographic and image distances. The best radius in this experiment is 113 m for OIF(1) and 130 m for OIF(2). The slope of the OIFs declines after the best radius.
Figure 10. Radius of range search vs. MRR
VI. CONCLUSION
We proposed a design for a geoconscious CBIR system based on a P2P mobile agent framework. Geoconscious CBIR will enhance the image retrieval accuracy for location-dependent photographs by limiting the geographic range for finding the image and using a combination of geographic and image distances. The P2P mobile agent framework will accelerate the performance and scalability of the retrieval.
We used the PIAX P2P mobile agent framework, as it adapts an overlay network and supports a structured P2P architecture. We proposed the design of different agents for this framework, as well as a protocol for constructing an image database and for CBIR. Furthermore, we showed the results of preliminary experiments using the implementation of an image database construction function and a geoconscious image search function.
We are currently combining the database construction and image feature calculation functions together, and developing an additional image retrieval function.
ACKNOWLEDGMENT
This work was supported by KAKENHI (22300035).
REFERENCES
[1] J. Kamahara, N. Tanaka, K. Hirayama, T. Nagamatsu, Y. Teranishi, and Y. Ariyoshi, “Design of GeoConscious P2P Content-based Image Retrieval,” Proc. 7th Intl. Conference on Broadband and Wireless Computing, Communication and Applications (BWCCA2012), Nov. 2012.
[2] Y. Teranishi, “PIAX: Toward a Framework for Sensor
Overlay Network,” Proc. 6th Annual IEEE Consumer Communications & Networking Conference (CCNC 2009 Workshops), Jan. 2009, pp. 1-5, http://piax.org/en/.
[3] J. Li and G. Zhang, “The State of the Art in Content-Based Image Retrieval in P2P Networks,” Proc. ACM 2nd International Conference on Internet Multimedia Computing and Service (ICIMCS’10), Dec. 2010, pp. 143-146, doi: 978-1-4503-0460-3/10/12.
[4] M. Batko, V. Dohnal, D. Novak, and J. Sedmidubsky, “MUFIN: A Multi-Feature Indexing Network,” Proc. IEEE 2nd International Workshop on Similarity Search and Applications (SISAP’09), Aug. 2009, pp. 158-159, doi: 10.1109/SISAP.2009.24.
[5] X. Meng, C. Feng, and Y. Wang, “Efficient Image Retrieval in P2P Using Distributed TS-SOM and Relevance Feedback,” Proc. International Symposium on Distributed Computing and Applications to Business; Engineering and Science (DCABES2007), Aug. 2007, pp. 1066-1071.
[6] J. Luo, D. Joshi, J. Yu, and A. Gallagher, “Geotagging in Multimedia and Computer Vision—A Survey,” Multimedia and Tools Application, vol. 51, no. 1, Jan. 2011, pp. 187-211, doi:10.1007/s11042-010-0623-y.
[7] Y. Li, D. J. Crandall, and D. P. Huttenlocher, “Landmark Classification in Large-scale Image Collections,” Proc. IEEE 12th International Conference on Computer Vision (ICCV 2009), Sept. 2009, pp. 1957-1964, doi: 10.1109/ICCV.2009.5459432.
[8] X. Li, C. Wu, C. Zach, S. Lazebnik, and J. M. Frahm, “Modeling and Recognition of Landmark Image Collections Using Iconic Scene Graphs,” Proc. ACM 10th European Conference on Computer Vision (ECCV ’08), Oct. 2008, pp.427-440, doi: 10.1007/978-3-540-88682-2_33.
[9] F. Aurenhammer, “Voronoi Diagrams — A Survey of a Fundamental Geometric Data Structure,” ACM Computing Surveys, vol. 23, no. 3, Sept. 1991, pp. 345-405, doi: 0360-0300/91/0900-0345.
[10] K. Hirayama, T. Matsui, and M. Yokoo, “Adaptive Price Update in Distributed Lagrangian Relaxation Protocol,” Proc. the 8th International Joint Conference on Autonomous Agents & Multi-Agent Systems (AAMAS-2009), May 2009, pp. 1033-1040.
[11] M. Ohnishi, Y. Minamoto, T. Eguchi, H. Kato, R. Nishide, and S. Ueshima, “Autonomous and Distributive Generation Algorithm of Delaunay Network for P2P Model Utilizing Node Location,” IPSJ Transaction on Database, vol. 47, Mar. 2006, pp. 51-64.
[12] J. Kamahara, T. Nagamatsu, and N. Tanaka, “Conjunctive Ranking Function using Geographic Distance and Image Distance for Geotagged Image Retrieval,” Proc. the ACM multimedia 2012 workshop on Geotagging and its Applications in Multimedia (GeoMM’12) , Nov. 2012, pp. 9-14.
[13] D. R. Radev, H. Qi, H. Wu, and W. Fan, “Evaluating Web-based Question Answering Systems.” Proc. of 3rd Int. Conf. on Language Resources and Evaluation (LREC 2002), May. 2002.
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0 50 100 150 200
MRR
Radius (m)
Geographic Distance Image Distance
OIF(1) OIF(2)
The 9th International Workshop on Mobile Peer-to-Peer Computing 2013, San Diego (18 March 2013)
77