Creating images the 2-D wayCreating images (3-D)Creating images (2-D + 3-D)Inserting objects into imagesInserting objects into images(2-D + 3-D)Inserting objects in imagesAlternative: Clip artCreating images (2-D)Photo-realistic CartoonExpensive and impractical Cheap and intuitive?Challengesobject orientationscene illuminationTHISThe use of dataSOMEQuickTime™ and aMPEG-4 Video decompressorare needed to see this picture.The Google modelDatabaseSort the objectsQuery Results2-D image vs 3-D sceneOutlineData source: LabelMeData organizationAnnotating the objectsCamera parametersHuman height distribution1.7 +/- 0.085 m(National Center for Health Statistics)Car height distribution1.5 +/- 0.19 m(automatically learned)Camera parametersObject heightsObject Estimated average height (m)Car 1.51Man 1.80Woman 1.67Parking meter 1.36Fire hydrant 0.87Estimated object heightsGeometry is not enoughIllumination contextet al.Illumination contextL*a*b*Automatic Photo PopupHoiem et al., SIGGRAPH ‘05Illumination nearest-neighborsOther criteria: local contextOther criteria: segmentationOther criteria: blurRecapLet’s insert an object!Seamset al.et al.Poisson blending: ideaWhy gradients? 1-D exampleRegular blendingbrightdark1-D exampleBlending derivativesOriginal signals DerivativesReintegration results1-D exampleGradient domainIntensity domain2-D: not so easyNon integrable: sum over a loop ≠ 0Actually happens all the time in practice2-D: some notation2-D: a (possible) solution?2-D: a (popular) solutionResults & limitationset al.Poisson blending: improvementset al.et al.Still not right!directionShadow transfer+=User interfaceQuickTime™ and aMPEG-4 Video decompressorare needed to see this picture.Street accidentBridgePaintingAlleyFailure casesShadow transferPorous objectsFailure casesBest matching objectsFailure casesNew data source: webcamsMatching across webcamsWebcam Clip ArtPros & consPros
View Full Document