Robust Method of Vote Aggregation and Proposition Verification for Invariant Local Features

0.54 MB·

by  Grzegorz Kurzejamski

Robust method of vote aggregation and proposition verification for
invariant local features
GrzegorzKurzejamski,JacekZawistowski,GrzegorzSarwas
LingaroSp.zo.o.
Puławska99a,02-595Warsaw,Poland
{grzegorz.kurzejamski,jzawisto,grzegorz.sarwas}@gmail.com
Keywords: ComputerVision,ImageAnalysis,MultipleObjectDetection,ObjectLocalization,PatternMatching
6 Abstract: Thispaperpresentsamethodforanalysisofthevotespacecreatedfromthelocalfeaturesextractionprocess
1 inamulti-detectionsystem. Themethodisopposedtotheclassicclusteringapproachandgivesahighlevel
0 of control over the clusters composition for further verification steps. Proposed method comprises of the
2 graphicalvotespacepresentation,thepropositiongeneration,thetwo-passiterativevoteaggregationandthe
cascadefiltersforverificationofthepropositions.Cascadefilterscontainalloftheminoralgorithmsneededfor
n
a effectiveobjectdetectionverification.Thenewapproachdoesnothavethedrawbacksoftheclassicclustering
J approachesandgivesasubstantialcontroloverprocessofdetection. Methodexhibitsanexceptionallyhigh
5 detectionrateinconjunctionwithalowfalsedetectionchanceincomparisontoalternativemethods.

]
V
C 1 INTRODUCTION ternandmatchthemagainstthemselves,usinganap-
. propriate metric. Common evaluations of such sys-
s
tems use brute KNN classification or its derivatives
c Objectdetectionbasedonlocalfeaturesiswellknown
[ as FLANN KNN search or BBF search. These give
in the computer vision field. Many different re-

good approximation of, theoretically ideal, results
1 searches brought about different features and meth-
with substantially lower computational cost. Some
v odsofsceneanalysisinsearchofaparticularobject.
of the applications make use of a LSH hashing but
1 Recently developed feature points are proven to be
8 wellsuitedforaspecificobjectdetection,ratherthan thisapproachdoesnotprovidepracticaldistancedata
7 needed for further analysis. Matching local features
a generalised object’s class identification. As many
0 provides set of correspondences, that can be filtered
applicationsoflocalfeatureshasbeenevaluated, the
0 later. Under the assumption, that the scene contains
ability to describe selective elements of rich graph-
.
1 oneornoinstanceoftheobject, correspondencecan
ics is the main purpose of invariant local features in
0 beputintotheobject-relatedornoise-relatedclass.To
manyfields. Amongstthemostpopularlocalfeatures
6 distinguish which class a particular correspondence
aree.g.SIFT,SURF,BRISK,FREAK,MSER.These
1
belongs to, few methods have been developed. Fre-
: are commonly described as feature points or feature
v quently used method for such purpose is RANSAC,
regions. They are easy to manipulate and to match.
i giving very good results, even for a high level of
X Useofsuchlocalfeaturesgivesnotonlytheabilityto
noise-related correspondences. After the classifica-
describetheobjectinmanyways,butalsotoachieve
r
tion of correspondences the model can be assumed
a invarianceforbasicobjectandimagetransformation,
and a homography can be calculated. Well designed
asskew,rotation,blur,noise. Invariantlocalfeatures
parameters and filters can lead to high detection rate
used in conjunction with invariant characteristic re-
andlowfalsedetectionchance. RANSACandquan-
giondetectors,suchasHarris-AffineorSIFTdetector,
titative analysis of correspondence data become in-
providedataforthoroughscene-objectanalysis,lead-
efficient, when a contribution of noise-related corre-
ing to detection of an object in the scene. Giving an
spondences in whole correspondence group grows.
appropriatesetoffeaturesonecandeterminetheex-
Because of that such approach is not sufficient for
actpositionoftheobjectinthescenewithaspecified
scenes, where the objects occupy small share of the
scale,rotationandevenminor,lineardeformations.
image. Methodsmentionedabovewillnotworkwell
The straightforward approach to a detection task
withmultipleobjectspresentintheimageaswell.
is to identify local features in the scene and the pat-
Systemsformulti-detectionpurposesincorporate 2 RELATED WORK
divide and conquer approach. Each correspondence
can be assigned to one of N+1 classes, where N is Localfeaturesintheimagecanbetrackedalongway
the number of objects in the scene. There is no in the literature. We present state-of-the-art feature
known, straightforward method of assigning corre- pointsextractinganddescribingmethods,thatcanbe
spondences. Most applications use correspondences usedinourmethod,andsimilarframeworksformulti-
as votes in a multidimensional space. Vote space object detection purposes developed through the last
canbeclusteredwithcommonclusteringalgorithms. years.
Eachclustercanbeprocessedwithasingle-objectde-
tection algorithm. Clustering approaches can be di- 2.1 Featurepoints
videdintotwogroups.Thefirstoneconsistsofsparse
clustering, where each cluster should contain all the
Our method should be used with conjunction with
needed vote data of a specific class. Such methods
scale-invariant and rotation-invariant features for the
show low detection rate because of a far from ideal
best results. Usage of local features lacking any of
clusteringprocessandahighlevelofparametrization.
this characteristics may come with a need for rejec-
Secondgroupisdenseclustering,whereclustersmay
tion of some parts of our method, but can be imple-
contain only small portion of a particular correspon-
mentednevertheless.
denceclass. Itsanalysisleadstocreatingorsupport-
Thebestknownfeaturepoints,uptothispoint,are
inghypothesisoftheobject’soccurrenceinthescene.
SIFTpointsdevelopedbyLowe (Lowe,1999),which
TheflagshipsinthatmatterareHough-likemethods.
became a model for various local features bench-
Such approaches show high detection rate but high
marks. TheclosestalternativetoSIFTisSURF(Bay
false positive rate as well. There are some works
et al., 2008), that comes with a lower dimension-
thattrytosegmentthescenewithknowncontext, as
ality and, in the result, a higher computing effi-
shown in the work of Iwanowski et al. (Iwanowski
ciency. There are also known attempts to incorpo-
etal.,2014).
rate additional enhancements into SIFT and SURF
Inourtestsbothclusteringapproacheslackedthe as PCA-SIFT (Ke and Sukthankar, 2004) or Affine-
abilitytoattainaveryhighdetectionratewithavery SIFT (Morel and Yu, 2009). SIFT and SURF and
low false positive rate at the same time. For our test its derivatives are computationally demanding dur-
cases a processing power is not a limitation and the ing matching process. In last years there has been
imagesareofaveryhighquality. Ourtestdatacon- big development in feature points based on binary
tains from zero to up to 100 objects per image and test pairs, that can be matched and described in a
presentsdifferentenvironmentalconditions. Mostof very fast manner. The flagships of this approach
the stat-of-the-art publications do not test detection are BRIEF (Calonder et al., 2010), ORB (Rublee
capabilities for such complex tasks. We found that et al., 2011), BRISK (Leutenegger et al., 2011) and
current approaches cannot maintain good detection FREAK (Alahi et al., 2012) features. Most of the
ratetofalsepositiverateratioonsatisfactorylevelin cited algorithms can be used to create dense and
manyreallifeapplications. highlydiscriminativevotingspace, whichholdssub-
stantialobjectcorrespondencedataneededtoaccom-
Thispaperpresentsthemethodofvotespaceanal-
plishmanyofthereal-worlddetectiontasks.
ysis,apartofinventionshownin(Kurzejamskietal.,
2014).Themethodcanbeadjustedtoavastvarietyof
2.2 Frameworks
objectdetectionpurposes,wheretheeffectivenessand
a low false positive rate is crucial. The method has
beendevelopedtoworkwellwithhugeamountoffea- There are few approaches to conduct multi-object
turedata,extractedfromhighqualityimages.Mostof multi-detection, meaning detecting multiple differ-
the algorithms used in the new approach come with ent objects on the scene, where any object can be
a logical justification. The new method uses the it- visible in multiple places. Viola and Jones (Viola
erative vote aggregation, starting from proposition’s and Jones, 2001) developed cascade of boosted fea-
positions. Propositions are generated from a graph- tures,thatcanefficientlydetectmultipleinstancesof
ical vote space analysis. Aggregated data undergoes the same object in a one pass of the detection pro-
analysisandfiltration. Wholeprocesshasatwo-pass cess. The method needs a time consuming, learn-
model,thatmakesthemethodrobusttosomespecific ing process with thousands of images. Method has
objectpositioninginthescene.Cascadeofaspecially been mostly tested on general objects, as people,
selectedsetoffilteralgorithmshasbeenutilizedtore- cars, faces. Most straightforward method for multi-
jectmostofthefalsepositivedetections. detectionisusingalloftheslidingwindowsasused,
forexample, inSarwas’andSkoneczny’swork(Sar- Algorithm 1: Vote Data, Vote Image and proposi-
was and Skoneczny, 2015). Most of them are un- tionscreation
Data: OriginalPatterns(OPT),SceneImage
fortunately computationally expensive. High effec-
(SCN)
tivenesscanbeachievedwithHistogramofOriented
Result: VoteDataandpropositionsfor
Gradients (Dalal and Triggs, 2005) and Deformable
object’scentresforeachpattern.
PartModels(Felzenszwalbetal.,2010). Thebiggest
drawbacksforourapplicationisthatDeformablePart 1 FeaturepointsextractiononOPTandSCN;
Models needs learning stage and Histogram of Ori- 2 foreachpatterninOPT do
ented Gradients is not rotation invariant. Blaschko 3 Findcorrespondences(COR)between
patternandSCNfeaturepoints;
and Lampert in (Blaschko and Lampert, 2008) uses
SVM to enhance the sliding window process. Ef- 4 foreachcorrespondenceinCORdo
ficient subwindows search has been used in (Lam- 5 Rejectifhaslowdistancevalue;
pert et al., 2008). In addition, branch-and-bound 6 Rejectifhashighhuedifferencevalue;
approaches, as in (Yeh et al., 2009), are promising 7 Calculateadjacencyvalue;
for multi-detection purposes with conjunction with 8 end
Bag-of-words descriptors. Lowe (Lowe, 2004) pro- 9 Creationofvotespace(VS)fromCOR;
posed generalized Hough Transform for clustering 10 Creationofvoteimage(VI)fromVS;
vote space with SIFT correspondence data. Authors 11 Searchforpropositions(PR)inVI;
of (Azad et al., 2009) created a 4D voting space 12 SortPRlist;
andusedcombinationofHough,RANSACandLeast 13 end
Squares Homography Estimation in order to detect
and accept potential object instances. Zickler in et
al. (Zickler and Efros, 2007) used angle differences 3.1 Voteimagecreation
criterion in addition to RANSAC mechanisms and
votenumberthreshold. Zickleretal.in(Zicklerand First part of our method is the vote space and vote
Veloso,2006)usedcustomprobabilisticmodelinad- imagecreation(lines9and10ofAlgorithm1). Vote
ditiontoHoughalgorithm. space consists of multiple dimensions: X, Y, Scale,
Rotation and Distance. Each vote contains specific
Inoursystem’sapplicationwecoulduseonlyone
X and Y position of the center of the object. The
genericpatternimageperobjectsowerejectedmost
Distance may be the result of using specific metric
ofthelearning-basedglobaldescriptors.
for particular feature points. For SIFT the standard
procedureistouseL2distanceforitsfeaturevector,
which contains gradient data in the area around the
characteristicpoint. Onemayusesomeadditionalin-
3 ALGORITHM
formationindistancecalculation,ascolordifference.
SomeonecanuserankingmethodastheLSHhashing
insteadoftheL2metricaswell.
The algorithm presented by authors is built upon Voteimageistheprojectionoftheadjacencydata
twomechanisms: thevotespacescreationandavote available in vote space onto X and Y dimensions.
aggregation for each of the vote spaces created. The Vote image has one intensity channel, created by
vote space is created for each pattern. Its adjacency normalizing adjacency sum cue. Another approach
dataisprojectedontothe(X,Y)plane,creatingvote would be to use the distance value instead of adja-
images (one for each vote space). The vote images cency as the main cue. We found L2 metric, as well
are analysed in search for object’s position proposi- as many other distance-based approaches, as insuffi-
tions. This mechanism is shown in the Algorithm 1. cient.
The aggregation process is performed for each vote Votesinthevotespacesarebuiltuponfilteredcor-
space and for its each proposition, starting from the respondencesets. Thedistancethresholdusedinline
propositionwiththehighestadjacencyvalue. Aggre- 5ofAlgorithm1wascalculatedas:
gation consists of two passes with slightly different
vote gathering approaches. The first pass is needed MIN(V)+MAX(V)
thr= , (1)
toestimatethedetectedobject’sareainthescene,so 2
thesecondaggregationpasswouldgatheronlyvotes where V is the votes group and the MIN and MAX
consideredtobefromthatparticularobject’sinstance. operatorsreturnthevalueofavotewithminimaland
The structure of each pass is presented in Algorithm maximaldistancevaluefromthegroup.Therejection
2. functionDispresentedinequation2.
Algorithm 2: Vote aggregation and detection ac- thevoteimageandcorrespondingpartofvotespace,
ceptance wherethepotentialobject’scenterislocated.Weused
Data: VI,VS,PR GoodFeaturesToTrackbyTomasiandShi(Shiand
Result: Occurrences(OCR)intheSCNfora Tomasi,1994)todetectmultiplelocalmaximasinthe
particularpattern vote image and used them as the propositions. The
1 foreachPropositioninPRdo number of propositions should be much higher than
2 GatherallvotesinlocalareafromVS; a number of objects in the image. It is trivial to set
3 Uniquefilteringforgatheredvotes(V); the Good Features To Track to find all the important
4 CascadefilteringforV; pointsintheimage,butitleadstogenerationofthou-
5 ifnotrejectedbycascadefilteringthen sands of propositions. Number of propositions will
6 Estimateobject’sarea; significantly impact the algorithm’s processing time,
7 GatherallvoteswithaFloodFill soitisnotpossibletoignoretheneedforatrade-off
algorithm; in detector’s parameter adjustment. For each propo-
8 UniquefilteringfornewV; sition’sX andY position theadjacency sumfor cor-
9 CascadefilteringfornewV; respondingvotesinvotespaceiscalculatedandused
10 ifnotrejectedbycascadefilteringthen forsortingpurposesinline12. Thehighestadjacency
11 Calculateobject’sarea; value proposition should be the first taken later into
12 CreateoccurrenceentryinOCR; a vote aggregation process. As the adjacency sum is
13 Eraseallvotedatainoccurrence’s proportional to the channel value in the vote image,
areainVSandVI; the cue for sorting stage is easy to compute. Sort-
14 else ing the propositions ensures, that the strongest vote
15 rejectproposition groupingwillbeprocessed first. Incaseofthe posi-
16 end tiveobjectrecognition,thevotedatacorrespondingto
17 else object’sdetectionareawillbeerasedfromvotespace
18 rejectproposition andimage.
19 end
20 end 3.2 Voteaggregation
Second part of our approach contains a vote aggre-
gation mechanism. Vote aggregation starts from a
(cid:26)
accept, dist(v)≤thr proposition’sposition, whichshouldbethecenterof
D(v)= . (2)
reject, dist(v)>thr a local vote grouping in the vote image. Data of the
votegroupingscansignificantlyvaryfordifferentob-
Wetransformdistancevalueintonormalizedadja-
jectinstancesinthescene. Thebestinstancescanbe
cencyvalueinrangefrom0to1(line7ofAlg. 1). 1
represented by hundreds of votes, when the weakest
indicatesperfectmatch. 0indicatesneartorejection
positive object response can be connected with only
difference between feature points. We transformed
a few. Generic clustering may ignore such clusters
distancevaluesintoadjacency(adj)valueswithaspe-
andmergeitwiththebiggerones. Genericclustering
cificfunction:
algorithmshasgenericparameters,thatarehardtoad-
justwithobject-orientedlogicorevenintuition.Some
(cid:18)dist(v)(cid:19)2
adj(v)=1− . (3) clusteringapproachestendtoclusteralltheavailable
thr vote data, even if the noise (false correspondences)
Adjacency values are gathered in a single chan- fillsmostofthevotespace.
nel, gray vote image. Vote image can be optionally We propose an iterative 2-pass vote aggregation
normalizedforvisualizationpurposes. Suchnormal- process for selective clustering purposes. In each
izedimagehasbeenshowninFigure1. Ifthefeature passtheuniquefilteringandthecascadefilteringtake
extraction and matching process are highly discrim- place,whichrejectfalsepositivedetections.Twopass
inative, the object instances in the scene should be design prevents situations in which aggregation area
recognizablebyahuman. Manualverificationofvote containsmultipleobjects.Passoneoftheaggregation
imagegivessomelevelofvaluableinsightintovotes collectsallthevotesinlocalareaofproposition’spo-
intersperse in the image and a level of a false votes sition(line2ofAlg. 2). Thesizeofalocalareamay
groupingsrecognizablebyahuman. be a function of a corresponding pattern size. After
Last step of Algorithm 1 is the search for propo- gatheringofallthevotesinthelocalarea,theunique
sitions in a vote image. Proposition is a point in filteringisperformedandtheresultinggroupofvotes
(a)Voteimage (b)Blurredandnormalizedvoteimage (c)Scenewithteadetections
Figure1:Sampleofvoteimagegeneratedwhilelocalizingred,herbalteacasing.
istestedwithacascadeoffilters(lines3and4ofAlg. Some of the false positive detections in our ex-
2). periments were initiated as a bunch of feature points
placedalongasimple,steepgradientsandedges. For
Inthesecondpassoftheprocesstheaggregationis
instance, when the scene presented product shelves,
conductedwithFloodFillalgorithm,startingfromthe
more than a half of false detections contained edge
proposition’s position (line 7 of Alg. 2). The Flood
of the shelf near the center and its vote data present
Fill range is limited to a scaled down object’s area.
mostlyalongtheshelf’sedge.
Such limitation can be constructed with a scale and
rotation estimation from the first pass of the aggre-
3.3 Cascadefiltering
gation. The limitation ensures, that the aggregation
process will not collect the votes from neighbouring
Cascade filtering (lines 4 and 9 of Alg. 2) is a pro-
object instances. Second pass of the algorithm con-
cess of validating vote group with a cascade of fil-
tainsuniquefilteringandcascadefilteringaswell,as
ters. Eachfiltercanacceptaggregatedvotesorreject
thevotecollectionmaybedifferentinthispass.
them. Anyrejectionwillresultindroppingtheaggre-
Foreachgroupofvotes,auniquefilteringshould
gation process and removing the processed proposi-
be performed in each pass (lines 3 and 8 of Alg.
tionfromthepropositionssortedqueue. Novotedata
2). Uniquefilteringpreservesonlyonevotewiththe
isremovedfromvotespaceorvoteimageinthatsit-
highest adjacency corresponding to the same feature
uation. If all the filters in first pass accepts the vote
pointinthepattern. Wecandoso, becausewewant
group, process may estimate the size and rotation of
the aggregated votes to be connected with only one
the object represented by the majority of votes (line
object. Ifmultiplevotesareconnectedwithonespe-
11ofAlg. 2).
cificfeatureinthepatternwecanassumethatonlythe
Cascadefilterscompriseof:(1)votecountthresh-
strongestvoteisn’tthenoise.
olding,(2)adjacencysumthresholding,(3)scalevari-
Most of the feature point detectors incorporate ancethresholding,(4)rotationvariancethresholding,
mechanismsofrejectingthepointslocatedalongthe (5) feature points binary test, (6) global normalised
edges.Unfortunately,thismechanismsworkonlyina luminance cross correlation thresholding. First pass
microscale. Inhighresolutionsomegraphicalstruc- of the vote aggregation uses filters: (1), (2), (3) and
tures, that for human seem as a straight edge, has (4). Second pass of the process uses filters: (3), (4),
a very complicated, uneven shape for characteristic (5)and(6).
points detector. Characteristic points located along Vote count thresholding is a simple filter, thresh-
the edges have similar features, so may be matched olding number of votes in the aggregated group.
with the same feature in the pattern. It leads to gen- Loweinhiswork(Lowe,2004)proposedgeneralised
eration of many false propositions, which can some- Houghtransformforobjectdetection. Inthismethod
timesbeacceptedbyacascadefilters. he has assumed that only three votes are enough to
identify the object. Unfortunately, such assumption process, where the theoretical object’s frame can be
leads to many false positive detections. Three local calculatedfromthedatafromthefirstpass.
featuresarenotenoughtodescribecomplex,generic
graphics. We tested vote count thresholding for val-
uesfrom3upto20. Wefound6astheoptimalvalue
4 EXPERIMENTS
forfilteringouttooweakresponses.Ifthevotegroup-
ingrepresentsrealobjectinstanceandhaslessthan6
votes,itmeansthattheprioralgorithmprocesseshas Ourtestingplatform,incorporatingmethoddescribed
tooloweffectiveness. inthispaper,hasbeendevelopedtosearchproductlo-
gosandcasingsonscenespresentingmarketshelves
Adjacencysumthresholdingrejectsallthegroups
and displays. The database used for the test for this
of votes with sum of adjacency values less than a
paper consists of 120 shelf photos taken in 12MPx
threshold value. This filter in certain circumstances
resolution and scaled down to 3MPx for testing pur-
can be used instead of the vote count thresholding.
poses. The pattern group consists of 60 generic pat-
Neverthelesstherejectiondatafromthesetwofilters
terns of logos and product wrappings. Each shelf
may give an insight into vote certainty levels of the
photowastestedwitheachoneofthepatterns,giving
detection. Evenhugevotegroupingswithmorethan
7200detectionprocesses. Thephotoscontainedusu-
100votesmayhaveaverylowadjacencysumvalue.
ally three classes of products so most of the patterns
Scale variance thresholding rejects all the groups
couldgenerateonlyfalsepositivedetections.Average
of votes with a scale value variance higher than the
numberofproductspresentedinthesceneswas23.6.
thresholdvalue.Onemayrebuildthisfilterintomech-
Patternswerescaledtohavethebiggersizebetween
anismseparatingnoisesignalfrompositivedetection
512 and 256 pixels. In the application of product
signalwithaGaussianmodel. Forourpurposessuch
search on the market shelves we describe high qual-
method is computationally too expensive. Simple
ityimagesasphotosbiggerthan2MPx,withamini-
variance thresholding rejects many false detections
mumoftenthousandspixelsforthesmallestsearched
andiseasytocompute.
objectandallofthelogostextreadableforahuman.
Rotation variance thresholding rejects all the
Ouraggregationapproachbasesitseffectivenessupon
groupsofvoteswithrotationvalue’svariancehigher
chosenlocalfeatures. WeusedSIFTimplementation
thanthethresholdvalue.Rotationvariancethreshold-
formainexperiments. Mainadvantageofourmethod
ingworksanalogicallytoscalevariancethresholding
laysinfilteringoutfalsedetectionsandprocessingall
but using the rotation values. A rotation variance is
possible occurrences. SIFT is a state-of-the-art fea-
notstraightforwardtocompute. Wesettwelvebuck-
tures detector and descriptor. Our test showed that
ets for rotation values and choose the three buckets
100% of the actual object instances were processed
withthehighestcountnumber.Itsresultantwastaken
through our cascade filtering with a proper proposi-
asanaveragerotation. Allthevalueshasbeenrotated
tion’slocation. That’sthankstodensepropositionde-
so the average rotation was assigned to 180 degrees.
tectionsandastraightforwardvoteimagecreation.
Thenthevariationinregardsto180degreeshasbeen
Detection effectiveness lays in proper vote group
computedandusedforthresholding.
filtering. Amountofpositivedetectionsrejecteddur-
Featurebinarytestusesfeaturepointscorrespon- ing cascade filtering results from all the computer
dence data preserved in each vote. We created mul- vision algorithms incorporated into detection system
tiple luminance binary tests for random feature pairs and can be hardly used to measure aggregation ef-
inthescene,whicharerepresentedbyvotesinaggre- fectiveness without proper comparisons with similar
gatedvotesgroup. Wecreatedidenticaltestsforcor- methodsinthesameapplicationfield. Falsedetection
responding feature points on the pattern side. Each rate yields more analytical data. We found no false
setofbinarytestsprovidedabinarystringthatcanbe positivedetectionsduringourtests,thatwerefaultof
comparedwithahammingdistance. Thenormalised insufficient description capability of feature descrip-
distancecanbethresholded. tor. Alloffalsedetectionsweretheresultoftooloose
Normalisedluminancecrosscorrelationisusedas parameters, that were needed for very high positive
alastfilter. Itneedstheexactobject’sgraphicspatch detectionrate. Neverthelesswecameacross203false
extractedfromthescene. It’scomputationallyexpen- detectionsin129of7200detectionprocesses,result-
sive,butcanfilteroutmanyfalsepositivedetections, inginmorethan1%(Table1)falsedetectionchance
that cannot be filtered by previous filters. The im- per detection process. This result seems low, but at
agesareresizedtothesizeof50x50pixelsbeforethe thesametimemeans66.2%chance,thatthefalsede-
calculation of the cross correlation. The filtering is tection will take place when looking for any product
conductedonlyinthesecondpassoftheaggregation instancefromourpatternsdatabase.
(a)Pattern
(b)Scene
Figure2:Sampleofdetectionresults
Ourmethodhasbeencomparedtothemethodus- cessed further to discriminate different variations of
ing the HOG descriptor. For the training stage we theproducts. Onecanusepartialpatternswithabag-
generated set of 60 derivative images for each pat- of-wordsapproachontopofouraggregationmethod
tern through small affine transformations. We used todoso.
all other patterns as a negative images. We used im-
plementation of HOG method, called Classifier Tool
For OpenCV and FANN (HOG, 2014). Our method
5 CONCLUSIONS
achieved only slightly better detection rate, but sig-
nificantlylowerchanceforfalsedetections.Theaver-
Inthispaperthemethodofvoteaggregationdesigned
agenumberoffalsedetectionswerealmosttwotimes
for use in multi-object multi-detection systems has
higherfortheHOGapproach(Table2).
beenintroduced. Aggregationprocessyieldspromis-
ingresultsintests,leadingtoanalysisofeachpoten-
Method DetectionRate FalseDetectionChance
tial object in the image. The unique filtering leaves
Ours 81.3% 1.79%
outmanyfalseobjectoccurrencepropositionsandthe
HOG 73.6% 21.42%
cascadefilteringrejectsmostofthefalsepositivede-
Table1: Detectionrateandfalsedetectionchanceforour tections,thatiscrucialforpresentedapplication.Sys-
tests. tem built upon the aggregation method can achieve
morethan80%detectionratewiththefalsedetection
chance below 2%. It is still far from industrial stan-
Method AverageNumberofFalseDetections
dards, but there are many places for improvement as
Ours 1.57
well.
HOG 3.18
Presentedmethodisdesignedtoanalyzeveryhigh
qualityimages. Imagesprocessedintestsweretaken
Table 2: Average number of false detections for process,
by a hand, resulting in high amount of blurred and
wherethefalsedetectionoccurred.
skewedvisualdata. Themethodofimageacquisition
During experiments with product casings we en- should be analyzed further. In future work we will
countered number of problems with association of incorporate estimates of the best parameters for pre-
detections to a specific result group. Some prod- sented method as well as solve simple parametriza-
ucts are very similar, with only slight local graphi- tion dependencies. We are going to test the system
caldifferences. Thisisparticularlytrueforthesame with a two-phase approach, where second phase of
brandwithdifferentaromasorcasingsizes. Figure2 thedetectionwouldusethepatternsextracteddirectly
presentsoneofsuchcases,whereteacasinghasiden- fromthescene. Thepatternsizehastoomuchimpact
tical logo for its few variations with one being visu- on the detection rate, as the feature points approach
allyverydifferentfromtheothers. Wedecidedtoin- works the best, when the objects in the scene and in
terpret only the visually off tea as a false detection. pattern images have the same size. We are going to
Inretailfieldtherestofthedetectionsshouldbepro- evaluateresizingoptionsforbetterdetectionresults.
ACKNOWLEDGEMENTS
Kurzejamski, G., Zawistowski, J., and Sarwas, G. (2014).
Apparatusandmethodformulti-objectdetectionina
This work was co-financed by the European Union digitalimage. EUPatent14461566.3.
withintheEuropeanRegionalDevelopmentFund. Lampert,C.H.,Blaschko,M.,andHofmann,T.(2008).Be-
yondslidingwindows:Objectlocalizationbyefficient
subwindow search. In Computer Vision and Pattern
Recognition,2008.CVPR2008.IEEEConferenceon,
REFERENCES pages1–8.
Leutenegger,S.,Chli,M.,andSiegwart,R.(2011).BRISK:
(2014). Classifier tool for opencv and fann v. 4.11.8. Binary robust invariant scalable keypoints. In Com-
http://classifieropencv.codeplex.com/. Accessed: puter Vision (ICCV), 2011 IEEE International Con-
2014-12-18. ferenceon,pages2548–2555.
Alahi,A.,Ortiz,R.,andVandergheynst,P.(2012).FREAK: Lowe, D. (1999). Object recognition from local scale-
Fast retina keypoint. In Computer Vision and Pat- invariantfeatures.InComputerVision,1999.ThePro-
ternRecognition(CVPR),2012IEEEConferenceon, ceedings of the Seventh IEEE International Confer-
pages510–517. enceon,volume2,pages1150–1157vol.2.
Azad, P., Asfour, T., and Dillmann, R. (2009). Combin- Lowe, D. (2004). Distinctive image features from scale-
ingHarrisinterestpointsandtheSIFTdescriptorfor invariant keypoints. volume 60, pages 91–110.
fast scale-invariant object recognition. In Intelligent KluwerAcademicPublishers.
RobotsandSystems,2009.IROS2009.IEEE/RSJIn- Morel,J.-M.andYu,G.(2009). ASIFT:Anewframework
ternationalConferenceon,pages4275–4280. forfullyaffineinvariantimagecomparison. SIAMJ.
Bay, H., Ess, A., Tuytelaars, T., and Gool, L. V. (2008). Img.Sci.,2(2):438–469.
Speeded-uprobustfeatures(SURF).ComputerVision
Rublee, E., Rabaud, V., Konolige, K., and Bradski, G.
and Image Understanding, 110(3):346 – 359. Simi-
(2011). ORB: An efficient alternative to SIFT or
larityMatchinginComputerVisionandMultimedia. SURF. In Computer Vision (ICCV), 2011 IEEE In-
Blaschko,M.andLampert,C.(2008). Learningtolocalize ternationalConferenceon,pages2564–2571.
objectswithstructuredoutputregression. InForsyth,
Sarwas, G.andSkoneczny, S.(2015). Objectlocalization
D.,Torr,P.,andZisserman,A.,editors,ComputerVi-
anddetectionusingvariancefilter. InChoras´, R.S.,
sion ECCV 2008, volume 5302 of Lecture Notes in
editor, Image Processing & Communications Chal-
ComputerScience, pages2–15.SpringerBerlinHei-
lenges6,volume313ofAdvancesinIntelligentSys-
delberg.
temsandComputing,pages195–202.SpringerInter-
Calonder, M., Lepetit, V., Strecha, C., andFua, P.(2010). nationalPublishing.
BRIEF: Binary robust independent elementary fea-
Shi, J.andTomasi, C.(1994). Goodfeaturestotrack. In
tures. In Daniilidis, K., Maragos, P., and Paragios,
ComputerVisionandPatternRecognition,1994.Pro-
N., editors, Computer Vision ECCV 2010, volume
ceedings CVPR ’94., 1994 IEEE Computer Society
6314 of Lecture Notes in Computer Science, pages
Conferenceon,pages593–600.
778–792.SpringerBerlinHeidelberg.
Viola,P.andJones,M.(2001).Rapidobjectdetectionusing
Dalal,N.andTriggs,B.(2005).Histogramsoforientedgra-
aboostedcascadeofsimplefeatures.InComputerVi-
dientsforhumandetection. InComputerVisionand
sionandPatternRecognition,2001.CVPR2001.Pro-
Pattern Recognition, 2005. CVPR 2005. IEEE Com-
ceedingsofthe2001IEEEComputerSocietyConfer-
puter Society Conference on, volume 1, pages 886–
enceon,volume1,pages511–518.
893vol.1.
Yeh,T.,Lee,J.,andDarrell,T.(2009). Fastconcurrentob-
Felzenszwalb, P., Girshick, R., McAllester, D., and Ra-
jectlocalizationandrecognition. InComputerVision
manan, D. (2010). Object detection with discrim-
and Pattern Recognition, 2009. CVPR 2009. IEEE
inatively trained part-based models. Pattern Analy-
Conferenceon,pages280–287.
sisandMachineIntelligence, IEEETransactionson,
Zickler,S.andEfros,A.(2007). Detectionofmultiplede-
32(9):1627–1645.
formable objects using PCA-SIFT. In Proceedings
Iwanowski, M., Zielin´ski, B., Sarwas, G., and Stygar, S.
ofthe22NdNationalConferenceonArtificialIntelli-
(2014). Identification of products on shop-racks by
gence-Volume2,AAAI’07,pages1127–1132.AAAI
morphologicalpreprocessingandfeature-baseddetec-
Press.
tion. In Chmielewski, L., Kozera, R., Shin, B.-S.,
andWojciechowski,K.,editors,ComputerVisionand Zickler,S.andVeloso,M.M.(2006). Detectionandlocal-
Graphics,volume8671ofLectureNotesinComputer izationofmultipleobjects.InHumanoidRobots,2006
Science, pages 286–293. Springer International Pub- 6th IEEE-RAS International Conference on, pages
20–25.
lishing.
Ke, Y. and Sukthankar, R. (2004). PCA-SIFT: a more
distinctiverepresentationforlocalimagedescriptors.
In Computer Vision and Pattern Recognition, 2004.
CVPR2004.Proceedingsofthe2004IEEEComputer
Society Conference on, volume 2, pages II–506–II–
513Vol.2.