+ All Categories
Home > Documents > Perceptually Consistent Example‐ Based Human Moon...

Perceptually Consistent Example‐ Based Human Moon...

Date post: 26-Sep-2020
Category:
Upload: others
View: 0 times
Download: 0 times
Share this document with a friend
22
Perceptually Consistent Example‐ Based Human Mo9on Retrieval Zhigang Deng, Qin Gu, Qing Li University of Houston Presented By: Daniel De Sousa
Transcript
Page 1: Perceptually Consistent Example‐ Based Human Moon Retrievaldavid/Classes/ICG/Talks/DeSousaCS637_Slides.pdf5 Run modified KMP‐matching algorithm to retrieve and rank matches Hierarchical

PerceptuallyConsistentExample‐BasedHumanMo9onRetrieval

ZhigangDeng,QinGu,QingLiUniversityofHouston

PresentedBy:DanielDeSousa

Page 2: Perceptually Consistent Example‐ Based Human Moon Retrievaldavid/Classes/ICG/Talks/DeSousaCS637_Slides.pdf5 Run modified KMP‐matching algorithm to retrieve and rank matches Hierarchical

GoalsandObjec9ves

•  Givenalargemo9ondatarepositoryandaquerymo9onsequence,retrievesimilarmo9onsequences

•  Ranktheminperceptuallyconsistentorder

•  Performthesearchefficiently

Page 3: Perceptually Consistent Example‐ Based Human Moon Retrievaldavid/Classes/ICG/Talks/DeSousaCS637_Slides.pdf5 Run modified KMP‐matching algorithm to retrieve and rank matches Hierarchical

MethodOverview

1  Subdividethehumanbodyintoameaningfulhierarchicalstructure

2  Segmenteachpartandpar99onoriginalsequencesinpart‐basedsequences,thennormalizeframelength

3  Extractmo9onpaSernsbydetec9ngandgroupingsimilarsegments

4  Segmentquerymo9onsequenceinpaSernlists

5  RunmodifiedKMP‐matchingalgorithmtoretrieveandrankmatches

Page 4: Perceptually Consistent Example‐ Based Human Moon Retrievaldavid/Classes/ICG/Talks/DeSousaCS637_Slides.pdf5 Run modified KMP‐matching algorithm to retrieve and rank matches Hierarchical

HierarchicalPar99on

Page 5: Perceptually Consistent Example‐ Based Human Moon Retrievaldavid/Classes/ICG/Talks/DeSousaCS637_Slides.pdf5 Run modified KMP‐matching algorithm to retrieve and rank matches Hierarchical

Representa9onBenefits

•  Hierarchyrepresenta9on– Controlgranularitytodescribevariouslevels– Mul9‐layerdescribescorrela9onamongvariousparts

•  Jointanglesvs.MarkerPosi9ons– Anglesdescribemo9onregardlessofboneorlimbsizes

Page 6: Perceptually Consistent Example‐ Based Human Moon Retrievaldavid/Classes/ICG/Talks/DeSousaCS637_Slides.pdf5 Run modified KMP‐matching algorithm to retrieve and rank matches Hierarchical

Mo9onSegmenta9on

•  Longsequencesofmo9ondatawithlargevaria9oninmovement

•  Automa9callydividesegmentsintomoremeaningfulsegments

•  Automatedmethodformo9onsegmenta9onfoundinBarbicetall.[2004]

Page 7: Perceptually Consistent Example‐ Based Human Moon Retrievaldavid/Classes/ICG/Talks/DeSousaCS637_Slides.pdf5 Run modified KMP‐matching algorithm to retrieve and rank matches Hierarchical

Adap9veSegmenta9onProcess

•  Segmentifdimensionalityofmo9onincreasessignificantly

•  Computeprobabilityforeachpointdependingonitsfitinthecurrentaccepteddistribu9on

•  Ateachpeakfollowingalocalminimum,newsegmentiflargerthanathreshold,R

•  Rvalueshigherforupperlevelsofhierarchy,duetomoredegreesoffreedom(performanceandstorageboost)

Page 8: Perceptually Consistent Example‐ Based Human Moon Retrievaldavid/Classes/ICG/Talks/DeSousaCS637_Slides.pdf5 Run modified KMP‐matching algorithm to retrieve and rank matches Hierarchical

SegmentNormaliza9on

•  Foreachbodypartinthehierarchy,sequencesaresegmented

•  Producesvaryingnumberofframesforthesame9merangeofmo9on

•  Normalizesallsequencestohavesamenumberofframes,usingCubicSplineinterpola9ontocreatenewframes

Page 9: Perceptually Consistent Example‐ Based Human Moon Retrievaldavid/Classes/ICG/Talks/DeSousaCS637_Slides.pdf5 Run modified KMP‐matching algorithm to retrieve and rank matches Hierarchical

Mo9onPaSernExtrac9on

•  Amo9onpaSernisarepresenta9vemo9onsegmentforhierarchicalbodypart

•  OncepaSernsareextractedtheycanbematchedandcompared

•  Adap9veK‐meansclusteringalgorithmgroupssimilarsegmentsintopaSerns

Page 10: Perceptually Consistent Example‐ Based Human Moon Retrievaldavid/Classes/ICG/Talks/DeSousaCS637_Slides.pdf5 Run modified KMP‐matching algorithm to retrieve and rank matches Hierarchical

Adap9veK‐meansclustering

•  Ini9alnumberofclusters(K)•  Applythistothemo9onsegmentsfrompart2

•  Iferrormetricislargerthenlimit,increasenumberofclustersandrepeat

•  BeSerresultswithmaximumclusteringerrorvs.averageclusteringerror(buttherewasaperformancepenalty)

Page 11: Perceptually Consistent Example‐ Based Human Moon Retrievaldavid/Classes/ICG/Talks/DeSousaCS637_Slides.pdf5 Run modified KMP‐matching algorithm to retrieve and rank matches Hierarchical

UnderlyingDataStructures

•  Theresultoftheseopera9onsproducethreedatastructures:– Kd‐treetostoremo9onpaSerns– Mo9onPaSernIndexLists•  18listsofpaSernindexes,oneforeachbodypart

– PaSernDissimilarityMap•  Matrixcomputedforallmo9onpaSernsbasedonEuclideandistance

Page 12: Perceptually Consistent Example‐ Based Human Moon Retrievaldavid/Classes/ICG/Talks/DeSousaCS637_Slides.pdf5 Run modified KMP‐matching algorithm to retrieve and rank matches Hierarchical

Mo9onRetrievalAlgorithm

Page 13: Perceptually Consistent Example‐ Based Human Moon Retrievaldavid/Classes/ICG/Talks/DeSousaCS637_Slides.pdf5 Run modified KMP‐matching algorithm to retrieve and rank matches Hierarchical

PaSernSimilarityComparison

•  GivenapaSernindexlist,compu9ngsimilarityisastringmatchingproblemwithsomecaveats– Notafixednumberofcharacters

– Cannotexpecttofindperfectmatches,waytoomanymo9onpaSerns

– Mustfactorin“quasi‐”matches

Page 14: Perceptually Consistent Example‐ Based Human Moon Retrievaldavid/Classes/ICG/Talks/DeSousaCS637_Slides.pdf5 Run modified KMP‐matching algorithm to retrieve and rank matches Hierarchical

Compu9ngSimilarity

•  ModifiedtheKMPstringmatchingalgorithm•  IftwopaSernindicessimilarityarewithinthresholdMaxTolerance,thenit’saquasi‐match

•  Ifnumberofconsecu9vequasi‐matchesisgreaterthanMinConNum,scoreisincreased

•  Ifnoquasi‐match,reducethescore

•  Matchingscorenormalizedbasedonlength

Page 15: Perceptually Consistent Example‐ Based Human Moon Retrievaldavid/Classes/ICG/Talks/DeSousaCS637_Slides.pdf5 Run modified KMP‐matching algorithm to retrieve and rank matches Hierarchical
Page 16: Perceptually Consistent Example‐ Based Human Moon Retrievaldavid/Classes/ICG/Talks/DeSousaCS637_Slides.pdf5 Run modified KMP‐matching algorithm to retrieve and rank matches Hierarchical

SimilarityPropaga9on

•  Wepropagatethesimilarityscoresupward

•  Severaladvantages

•  Method:Sprfl=½xSrf+½xSrlSurfl=αxSrfl+(1‐α)xSprfl

Page 17: Perceptually Consistent Example‐ Based Human Moon Retrievaldavid/Classes/ICG/Talks/DeSousaCS637_Slides.pdf5 Run modified KMP‐matching algorithm to retrieve and rank matches Hierarchical

ResultsandComparison

Page 18: Perceptually Consistent Example‐ Based Human Moon Retrievaldavid/Classes/ICG/Talks/DeSousaCS637_Slides.pdf5 Run modified KMP‐matching algorithm to retrieve and rank matches Hierarchical

UserStudy

•  Askuserstoratemo9onretrievalssimilarityfrom1to10

Page 19: Perceptually Consistent Example‐ Based Human Moon Retrievaldavid/Classes/ICG/Talks/DeSousaCS637_Slides.pdf5 Run modified KMP‐matching algorithm to retrieve and rank matches Hierarchical
Page 20: Perceptually Consistent Example‐ Based Human Moon Retrievaldavid/Classes/ICG/Talks/DeSousaCS637_Slides.pdf5 Run modified KMP‐matching algorithm to retrieve and rank matches Hierarchical

Correla9onBetweenUserandNumericalAnalysis

Page 21: Perceptually Consistent Example‐ Based Human Moon Retrievaldavid/Classes/ICG/Talks/DeSousaCS637_Slides.pdf5 Run modified KMP‐matching algorithm to retrieve and rank matches Hierarchical

Conclusion

•  Par9cularlygoodatiden9fyingcomplexmo9on

•  Matchesorbeatscurrentmethodsinnumericalandperceptualra9ngs

•  However,cannotspecifytherootpathmo9on– Lenvs.rightmovement

Page 22: Perceptually Consistent Example‐ Based Human Moon Retrievaldavid/Classes/ICG/Talks/DeSousaCS637_Slides.pdf5 Run modified KMP‐matching algorithm to retrieve and rank matches Hierarchical

THANKYOU


Recommended