Leveraging Human Perception and Computer Vision Algorithms for Interactive Fine-Grained Visual

131 Pages·2014·32.1 MB·English

UC San Diego
UC San Diego Electronic Theses and Dissertations
Title
Leveraging Human Perception and Computer Vision Algorithms for Interactive Fine-Grained
Visual Categorization /
Permalink
https://escholarship.org/uc/item/5z2523gj
Author
Wah, Catherine Lih-Lian
Publication Date
2014

Peer reviewed|Thesis/dissertation
eScholarship.org Powered by the California Digital Library
University of California
UNIVERSITYOFCALIFORNIA,SANDIEGO
LeveragingHumanPerceptionandComputerVisionAlgorithmsforInteractive
Fine-GrainedVisualCategorization
Adissertationsubmittedinpartialsatisfactionofthe
requirementsforthedegreeofDoctorofPhilosophy
in
ComputerScience
by
CatherineLih-LianWah
Committeeincharge:
ProfessorSergeBelongie,Chair
ProfessorDavidKriegman
ProfessorGertLanckriet
ProfessorLawrenceSaul
ProfessorNunoVasconcelos
2014
Copyright
CatherineLih-LianWah,2014
Allrightsreserved.
TheDissertationofCatherineLih-LianWahisapprovedandisacceptable
inqualityandformforpublicationonmicrofilmandelectronically:
Chair
UniversityofCalifornia,SanDiego
2014
iii
DEDICATION
Tomyparentsandtwinsister,whocontinuetosupportandinspiremeeveryday.
iv
EPIGRAPH
Donotfeargoingforwardslowly;fearonlytostandstill.
Chineseproverb
Nothingisdifficult,onlyunfamiliar.
Unknown
GOD:IownyoulikeIownthecaves.
THEOCEAN:Notachance. Nocomparison.
GOD:Imadeyou. Icouldtameyou.
THEOCEAN:Atonetime,maybe. Butnotnow.
GOD:Iwillcometoyou,freezeyou,breakyou.
THEOCEAN:Iwillspreadmyselflikewings. Iamabilliontinyfeathers. Youhaveno
ideawhat’shappenedtome.
DaveEggers
v
TABLEOFCONTENTS
SignaturePage........................................................ iii
Dedication ........................................................... iv
Epigraph............................................................. v
TableofContents ..................................................... vi
ListofFigures ........................................................ viii
Acknowledgements.................................................... xi
Vita ................................................................. xv
AbstractoftheDissertation ............................................. xvii
Chapter1 Introduction ............................................... 1
Chapter2 BackgroundandRelatedWork................................ 4
2.1 Visipedia..................................................... 4
2.2 RelatedWork ................................................. 4
2.2.1 Fine-GrainedCategorization.............................. 4
2.2.2 Human-In-The-LoopMethods ............................ 6
2.2.3 ActiveClassification .................................... 7
Chapter3 InteractiveCategorizationwithPartsandAttributes .............. 9
3.1 Introduction .................................................. 9
3.2 RelatedWork ................................................. 10
3.3 VisualRecognitionwithHumansintheLoop ...................... 11
3.3.1 AlgorithmsandFramework............................... 14
3.3.2 IncorporatingComputerVision ........................... 15
3.3.3 ModelingUserResponses................................ 16
3.4 ExtensiontoPart-BasedModels ................................. 18
3.5 DatasetsandImplementationDetails ............................. 32
3.6 Experiments .................................................. 35
3.6.1 MeasuringPerformance.................................. 35
3.6.2 UsingBinaryAttributeQuestions ......................... 36
3.6.3 1-vs-allVs. Attribute-BasedClassification .................. 40
3.6.4 UsingPartandAttributeQuestions ........................ 41
3.7 Conclusion ................................................... 45
Chapter4 InteractiveCategorizationwithSimilarityLearning .............. 47
vi
4.1 Introduction .................................................. 47
4.2 RelatedWork ................................................. 50
4.3 PerceptualSimilarityMetricsforInteractiveCategorization .......... 51
4.3.1 MethodsandFramework................................. 52
4.3.2 IncorporatingComputerVision ........................... 58
4.4 ExtensiontoMultipleLocalizedPerceptualMetrics ................. 60
4.5 DatasetandImplementationDetails .............................. 65
4.6 Experiments .................................................. 68
4.6.1 EmbeddingGeneration .................................. 68
4.6.2 UsingNonlocalizedSimilarityMetrics ..................... 70
4.6.3 UsingMultipleLocalizedSimilarityMetrics ................ 79
4.7 HumanPerceptionofSimilarity.................................. 84
4.8 Conclusion ................................................... 87
Chapter5 Conclusion ................................................ 90
5.1 FinalThoughts................................................ 90
5.2 FutureDirections.............................................. 91
AppendixACUB-200-2011Dataset...................................... 92
A.1 Introduction .................................................. 92
A.2 DatasetSpecificationandCollection.............................. 93
A.3 Applications .................................................. 94
A.4 BenchmarksandBaselineExperiments ........................... 95
Bibliography ......................................................... 104
vii
LISTOFFIGURES
Figure2.1. Visipedia. ............................................... 5
Figure3.1. ScreencaptureofaniPadappforbirdspeciesrecognition. ... 10
Figure3.2. Examplesofclassificationproblems........................ 11
Figure3.3. Examplesofthevisual20questionsgame .................. 12
Figure3.4. Visualizationofthebasicalgorithmflow. ................... 14
Figure3.5. Examplesofuserresponses ............................... 17
Figure3.6. Interactivevisualrecognitionwithlocalization. ............. 19
Figure3.7. ProbabilisticModel. ..................................... 23
Figure3.8. FullyAutomatedPartDetectionResults.................... 25
Figure3.9. Userinterfaceforpartlocationsinput. ..................... 26
Figure3.10. Comparing part prediction accuracy for humans and comput-
ers ..................................................... 28
Figure3.11. Poseclusters............................................. 35
Figure3.12. DifferentModelsofUserResponses........................ 37
Figure3.13. PerformanceonBirds-200whenusingcomputervision ...... 38
Figure3.14. Examples where computer vision and user responses work to-
gether .................................................. 39
Figure3.15. Imagesthataremisclassifiedbyoursystem................. 40
Figure3.16. PerformanceonAnimalsWithAttributes .................. 41
Figure3.17. AttributeandPartQuestions. ............................. 42
Figure3.18. InteractiveClassificationUsingPartandAttributeQuestions. 43
Figure3.19. Examplesofthebehaviorofoursystem..................... 44
Figure4.1. Similaritymetricsforinteractivecategorization. ............ 48
viii
Figure4.2. InterfaceforCollectingSimilarityComparisons. ............ 49
Figure4.3. Localized Similarity Comparisons for Interactive Categoriza-
tion..................................................... 62
Figure4.4. DiscoveringDiscriminativeRegions. ....................... 63
Figure4.5. ComparingLocalizedandNonlocalizedComparisons. ....... 65
Figure4.6. Discriminative Regions. The 106 discovered discriminative re-
gions. Weselect5touseinourexperiments. .................. 67
Figure4.7. EmbeddingGeneralizationError. ......................... 70
Figure4.8. NonlocalizedSimilarityEmbedding. ....................... 71
Figure4.9. LocalizedSimilarityEmbeddingforRegion1................ 72
Figure4.10. LocalizedSimilarityEmbeddingforRegion13............... 73
Figure4.11. LocalizedsimilarityembeddingforRegion21. .............. 74
Figure4.12. LocalizedsimilarityembeddingforRegion23. .............. 75
Figure4.13. LocalizedsimilarityembeddingforRegion39. .............. 76
Figure4.14. Test-TimeInterface. Anexampleofatest-timeuserinterfacefor
ourinteractiveclassificationsystem. ......................... 77
Figure4.15. ObservingDeterministicUsers. ........................... 77
Figure4.16. ObservingSimulatedNoisyUsers. ......................... 78
Figure4.17. QualitativeResultsforNonlocalizedMetrics................ 80
Figure4.18. QualitativeResultsforLocalizedMetrics. .................. 81
Figure4.19. InteractiveCategorizationResults. ........................ 82
Figure4.20. QuestionDistribution..................................... 85
Figure4.21. Comparing human perception of nonlocalized vs. localized
similarity................................................ 86
Figure4.22. ObservingAMTworkerbehavior. ......................... 88
ix