low bit rate mobile visual search
TRANSCRIPT
-
8/4/2019 Low Bit Rate Mobile Visual Search
1/59
InstituteofDigitalMedia,TheSchoolofCS&EE,PekingUniversity,ChinaAug.18,2011,NTU,Singapore
-
8/4/2019 Low Bit Rate Mobile Visual Search
2/59
2
-
8/4/2019 Low Bit Rate Mobile Visual Search
3/59
3
-
8/4/2019 Low Bit Rate Mobile Visual Search
4/59
ProductPackaging
LandmarksLogos
Faces
Cars Pets
UnpackagedProductsigidTexturedArticulateContourDefined
Momentoftruthforcomputervision Instancematching,notclassrecognition
4
-
8/4/2019 Low Bit Rate Mobile Visual Search
5/59
PervasiveMobileVisualSearch
DevelopedSystems
Mobot
NevenVision
Ideixis
Snaptell
PointandFound
GoogleGoggle
-
8/4/2019 Low Bit Rate Mobile Visual Search
6/59
SideInformationisuseful
formobilevisualsearch
VisualSearch
SideInformation
Mobile
VisualSearch
Highefficient
visualvocabularycoding
Desirablesearch
precision&fastspeedover
extremelylargescaledatasets
Goodscalabilityinterms
ofupstreamquery
coding
complexity
-
8/4/2019 Low Bit Rate Mobile Visual Search
7/59
PervasiveSide(Context)Information
GPS AccessPoints RFIDTags
VisuallyEasilyRecognizedBarcode,
LabelsorLogos
-
8/4/2019 Low Bit Rate Mobile Visual Search
8/59
Amobilephotoqueryhastobetransmittedfromamobile
devicetoaserver
Oftenoverarelativelyslowwirelesslink3G
Thequalityofuserexperienceheavilydependsonhowmuchinformationhastobetransmitted,especiallyinthe
scenarioofrealityaugmentation
8
-
8/4/2019 Low Bit Rate Mobile Visual Search
9/59
Compactlocalfeaturedescriptors e.g.CHoG Stanford,DAISYEFPL,PCASIFTDescriptorsCMU
Compactimagesignatures e.g.miniBoW,AggregateLocalDescriptorsINTRIA
Previousworksstilldemand2KBto4KBvisualdescriptorsperqueryimageunsuitableforrealityaugmentation
Withinanunstablewirelessnetworke.g.3Gnetwork,the
deliverycoststilloccasionallydelaysthequery9
-
8/4/2019 Low Bit Rate Mobile Visual Search
10/59
TransformCodingofSIFT/SURFdescriptors
Chandrasekharetal.,VCIP09 Directcompressionoforientedimagepatch
M.Makaretal.,ICASSP09
Descriptordesignedforcompressibility:CHoGChandrasekharetal.,CVPR09
TreeHistogramCoding
Chenetal.,DCC09 CompressionofSpatialLayoutofLocalFeatures
Tsaietal.,Mobimedia10
10
-
8/4/2019 Low Bit Rate Mobile Visual Search
11/59
CompactDescriptorforVisualSearch AnongoingMPEGADHOCGroup
Standardschedule 2010,AprilInitializetheCDVSADHOCGroup
2011,July,FinalCFP
Evaluationset Mobilequery,especiallylandmarkqueries
MajorRequirementofCompactDescriptors
Compactness
Discriminability11
-
8/4/2019 Low Bit Rate Mobile Visual Search
12/59
12
-
8/4/2019 Low Bit Rate Mobile Visual Search
13/59
Allstateof theartcompactdescriptorsaredesigned
onlybasedonvisualcontentstatistics
Incorporatethecheaplyavailablemobilecontexte.g.GPS
orbasestationtagstosupervise thedescriptordesign13
-
8/4/2019 Low Bit Rate Mobile Visual Search
14/59
Effectiveandefficient,lowbitratemobilevisualsearch
sayhundredsofbitspervisualquery
Coding/Compressionofanoriginallyhighdimensional
imagesignaturese.g.BoW
14
M islearntfromthemobilecontext
-
8/4/2019 Low Bit Rate Mobile Visual Search
15/59
-
8/4/2019 Low Bit Rate Mobile Visual Search
16/59
Userentersagivenregion
Themobileenddownloadsorprestoretheregion
specificdescriptionfunction
Themobileusertakesaquery
ThemobileendextractstheinitialBoW,compressesit
intoaregionspecificcompactdescriptor 16
-
8/4/2019 Low Bit Rate Mobile Visual Search
17/59
Geographicalregiondivision
Learningcompactdescriptorwithineachregion
-
8/4/2019 Low Bit Rate Mobile Visual Search
18/59
18
-
8/4/2019 Low Bit Rate Mobile Visual Search
19/59
CrawledfromFlickrandPanoramio websites
Beijing
NewYorkCity
Barcelona
Singapore
Florence
ThegeographicaldistributionofFlickerandPanoramioPhotosinBeijing 19
-
8/4/2019 Low Bit Rate Mobile Visual Search
20/59
20
Notations:intotalmregionscurrentgeotaggedphotoxregionassignmentyi withim
-
8/4/2019 Low Bit Rate Mobile Visual Search
21/59
21
ExpectationStep
MaximizationStep
-
8/4/2019 Low Bit Rate Mobile Visual Search
22/59
22
ThegeographicaldistributionofFlickerand
PanoramioPhotosinBeijingCity
ThevisualawarepartitionofBeijing
intogeographicalregions.Different
colorsdenotedifferentclusters.
-
8/4/2019 Low Bit Rate Mobile Visual Search
23/59
23
-
8/4/2019 Low Bit Rate Mobile Visual Search
24/59
24
Forthequerybelongingregion,compactcodewordsarelearnttotransmit
Thelearntcompactcodebookandtheextracteddescriptorsinexemplarqueries.
Left:thecompactcodebookinthequerysassignedregion;Middle:thequery,
wherecolorhighlightsdenotethedetecteddescriptorsonthequery;Right:the
transmittedwords.Weonlytransmittheiroccurrenceindexinpractice.
-
8/4/2019 Low Bit Rate Mobile Visual Search
25/59
Learningacompressionmatrix
MMK
MK
ThecompactcodebookU isgeneratedviaUMTV UsescalablevocabularytreeSVTtobuildtheinitialhigh
dimvocabulary Hierarchicalquantizationoflocalfeatures
AgoodcodebookU shouldminimizetherankingloss
25
where istherankingpositionweightofIxwithrespecttoIq
,suchthatahigherrankcorrespondstoalargeweight.
-
8/4/2019 Low Bit Rate Mobile Visual Search
26/59
26
-
8/4/2019 Low Bit Rate Mobile Visual Search
27/59
withsampleloss,Boosting
ErrorWeighting Ranking DecodedSignature
27
Given aregion,aimtominimizetheoverallcost:
-
8/4/2019 Low Bit Rate Mobile Visual Search
28/59
Findtheithbestcodeword
Errorweightingupdate
Compressionfunctionupdate
28
-
8/4/2019 Low Bit Rate Mobile Visual Search
29/59
Formoredetails,pls refertoLocationDiscriminativeVocabularyCodingfor
MobileLandmarkSearch,Int.JournalofComputerVision,InPress
-
8/4/2019 Low Bit Rate Mobile Visual Search
30/59
30
-
8/4/2019 Low Bit Rate Mobile Visual Search
31/59
Achickenandeggproblem
Solvedviaaniterativecooptimization
31
-
8/4/2019 Low Bit Rate Mobile Visual Search
32/59
32
Codinglength ofregioni Themaximalcodinglength ofallregions
-
8/4/2019 Low Bit Rate Mobile Visual Search
33/59
CompactnessLearningProcessoverBeijingCity
33ThegeographicalvisualizationofthedescriptorcompactnessinBeijingcitythroughiterativecooptimizationT1to20,whichwouldbetalkedlater
Wenormalizetheminvs.max
ratioofdescriptorlengthsand
maptheratiotothesaturationofredcolor.Thegreenpoints
denotethedistributionofgeo
taggedphotos.Ingeneral,less
saturatedmapcorrespondsto
moreoptimaldescriptors.
-
8/4/2019 Low Bit Rate Mobile Visual Search
34/59
34
-
8/4/2019 Low Bit Rate Mobile Visual Search
35/59
Over1MgeographicaltaggedphotosfromFlickrandPanoramio websites
Beijing,NewYorkCity,Barcelona,Singapore,etc..
Fromthegeographicalmapofeachcity,wechoose30
mostdenseregionsand30randomregions Askvolunteerstomanuallyidentifyoneormoredominantviews.
Allnearduplicatedlandmarkphotostoagivenviewarelabeledin
itsbelongingandnearbyregions
35
-
8/4/2019 Low Bit Rate Mobile Visual Search
36/59
Nister etal.2006
Jegou etal.,2010
Chandrasekharetal.,2009a
Chenetal.,2009
Ouralternativeapproach36
-
8/4/2019 Low Bit Rate Mobile Visual Search
37/59
-
8/4/2019 Low Bit Rate Mobile Visual Search
38/59
-
8/4/2019 Low Bit Rate Mobile Visual Search
39/59
39
mAP withrespecttodifferentregions
mAP variancesindifferentregions,wedrawtwodimensionallatticesto
dividedregionswithrespecttodifferentimagevolumesanddescriptor
bits,thenaveragethemAP forregionsfallingintoeachlattice.
-
8/4/2019 Low Bit Rate Mobile Visual Search
40/59
40
-
8/4/2019 Low Bit Rate Mobile Visual Search
41/59
41
-
8/4/2019 Low Bit Rate Mobile Visual Search
42/59
-
8/4/2019 Low Bit Rate Mobile Visual Search
43/59
43
-
8/4/2019 Low Bit Rate Mobile Visual Search
44/59
ensureinteroperabilityof
visualsearchapplications
anddatabases,
enablehighlevelof
performanceof
implementations
conformant
tothestandard,
simplifydesignofdescriptor
extractionandmatchingfor
visualsearchapplications,
enablehardwaresupportfordescriptorextractionand
matchinginmobiledevices,
reduceloadonwireless
networkscarryingvisual
searchrelatedinformation.
-
8/4/2019 Low Bit Rate Mobile Visual Search
45/59
Meeting# Date Action97 July1822,2011 FinalCfPissued
98 Nov.28Dec.02,2011 Initialevaluationofproposalsandassignmentofcrosschecks
99 Feb.0610,2012 WD1
101 July1620,2012 CD
103 Jan.2125,2013 DIS
105 July,2013 FDIS
-
8/4/2019 Low Bit Rate Mobile Visual Search
46/59
Requirement Description Fulfillmentinformation
Sufficiency Descriptorsshallbeselfcontained,inthesensethatno
otherdataarenecessaryformatching
Y/N
Formatindependence
Descriptorsshallbeindependentoftheimageformat Y/N
Robustness Highmatchingaccuracyshallbeachievedatleastfor
imagesoftexturedrigidobjects,landmarks,andprinted
documents.Thematchingaccuracyshallberobustto
changesinvantagepoint,cameraparameters,lighting
conditions,aswellasinthepresenceofpartial
occlusions
*Retrievalandpair
wisematching
accuracyresults
obtainedfordifferent
descriptorlengths.
Compactness Shallminimizelengths/sizeofimagedescriptors
Scalability 1. Shall allow adaptation of descriptor lengths to
support the required performance level and
database size.2. Shall enable design of webscale visual search
applications and databases.
Rangeofdescriptor
lengthsanddatabase
sizes
supported.
Extractioncomplexity
Shallallowdescriptorextractionwithlowcomplexity(in
termsofmemoryandcomputation)
*Timinginformation
fordescriptor
extractionopertations.
Matchingcomplexity
1. Shall allow matching of descriptors with low
complexity (in terms of memory and
computation).
2. If decoding of descriptors is required formatching, such decoding shall also be possible
with low complexity.
*Timinginformation
forretreivalandpair
wisematching
opertations.
Localization 1. Shall support visual search algorithms that
identify and localize matching regions of the
query image and the database image
2. Shall support visual search algorithms that
provide an estimate of a geometric
transformation between matching regions of the
query image and the database image
*Localizationaccuracy
results.
(*)resultsobtainedusingdatabasesandtestproceduresspecifiedintheevaluationframework[1].
-
8/4/2019 Low Bit Rate Mobile Visual Search
47/59
The Framework of Multichannel
-
8/4/2019 Low Bit Rate Mobile Visual Search
48/59
TheFrameworkofMulti channel
CompactVisualDescriptor(MCVD)
Theencodedcompactdescriptorincludestwoparts:
ChannelIdentification +ChannelDependentVisualDescriptorNote:
A
channel
refers
to
subdivide
(partition)
the
reference
database.
Within
each
channel,
wemay generateachanneldependent(extremely)compactdescriptor.
The Framework of Multichannel
-
8/4/2019 Low Bit Rate Mobile Visual Search
49/59
Thechannellearningcanbebasedoncontextualtags(e.g.GPSorRFID)
orcanbesolelyonthevisualstatistics.
TheFrameworkofMulti channel
CompactVisualDescriptor(MCVD)
Wh d di i k l
-
8/4/2019 Low Bit Rate Mobile Visual Search
50/59
GPS
Access
Points
RFIDTags
StandardizationEfforts
VisuallyRecognized
Barcode,Logos,etc.
Wherestandardizationtakesplace
J ifi i f Ch l i MCVD
-
8/4/2019 Low Bit Rate Mobile Visual Search
51/59
ChannelNum=1
Channel1Model Channel2Model ChannelKModel
ChannelDivision
JustificationofChannelinMCVD
J tifi ti f Ch l i MCVD
-
8/4/2019 Low Bit Rate Mobile Visual Search
52/59
ChannelNum=1
Channel1Model Channel2Model ChannelKModel
ChannelDivision BoostingCompactCodebook
Channel1
Compact
Codebook
Channel2
Compact
Codebook
JustificationofChannelinMCVD
-
8/4/2019 Low Bit Rate Mobile Visual Search
53/59
53
TwotypicalMCVDcaseswithorwithoutcontextualinformation,with
comparisontothestateoftheart,overmillionscalelandmarkdataset.
-
8/4/2019 Low Bit Rate Mobile Visual Search
54/59
54
ComparisonofmAP withrespecttotheupstreamquerybitratefor
product/CD/bookcoverBenchmarkDatabases(publiclyavailable).
-
8/4/2019 Low Bit Rate Mobile Visual Search
55/59
55
Search
Battery
Bandwidth
-
8/4/2019 Low Bit Rate Mobile Visual Search
56/59
56
Phone Battery:
Voltage: 4.0 V, Capability:
1400 mAH (or 20.2K Joules)
Sending Images
20.2K Joules/52.4 Joules
~ 385 queries in total
Sending MCVD:
20.2K Joules/8.1Joules
~ 2494 queries in total
Averageenergyconsumptioncomparisonthroughthe3Gwireless
link,betweentransmittingtheentirequeryimageandtheextracting
andtransmittingofMCVDandothercompactdescriptors.
-
8/4/2019 Low Bit Rate Mobile Visual Search
57/59
57
O l t
-
8/4/2019 Low Bit Rate Mobile Visual Search
58/59
Ourrelevantpapers
SelectedPapers LocationDiscriminativeVocabularyCodingforMobileLandmarkSearch,Int.JournalofComputerVision,InPress LearningCompactVisualDescriptorforLowBitRateMobileLandmark
Search,IJCAI2011,Barcelona,Spain,Jul.2011 TowardsLowBitRateMobileVisualSearchwithMultipleChannel
Coding,ACMMM2011,Arizona,USA,Nov.2011 MPEGInputContributions
CompactDescriptorsforVisualSearch,m18542,MPEG94th PKUBench:AContextualRichBenchmarkforMobileVisualSearch,
m19188,MPEG95th MultipleChannelCompactVisualDescriptorwithAdaptiveChannel
Learning,m19985,MPEG96th TopiclevelSamplingTowardsOptimizedLocalitySensitiveVocabulary
Coding,m21199,MPEG97th
Bandwidth
-
8/4/2019 Low Bit Rate Mobile Visual Search
59/59
Thanks!
Q & A
Visual
Search
Side
Information
MobileVisual
Search
Search
Battery