SURF：Speed Up Robust Feature.ppt

《SURF：Speed Up Robust Feature.ppt》由会员上传分享，免费在线阅读，更多相关内容在教育资源-天天文库。

1、1SURF:SpeededUpRobustFeatures,ECCV2006.HerbertBay,TinneTuytelaars,andLucVanGoolGroupMeetingPresentedbyWyman10/14/20062BackgroundLocalinvariantInterestpointdetector-descriptorForfindingcorrespondencesbetweentwoimagesofthesamesceneorobjectManyapplicati

2、ons,including3Dreconstruction,imageretrievalandobjectrecognitionSIFTisoneofthebestbutslowImageofsize1000x700describedinaround6seconds(actualcostdependsonthe#featuresgenerated,4000inthiscase)128-Dfeaturevectors3MotivationFastinterestpointdetectionDist

3、inctiveinterestpointdescriptionSpeeded-updescriptormatchingInvarianttocommonimagetransformations:ImagerotationScalechangesIlluminationchangeSmallchangeinViewpoint4MethodologyUsingintegralimagesformajorspeedupIntegralImage(summedareatables)isaninterme

4、diaterepresentationfortheimageandcontainsthesumofgrayscalepixelvaluesofimageSecondorderderivativeandHaar-waveletresponseCostfouradditionsoperationonly5DetectionHessian-basedinterestpointlocalizationLxx(x,y,σ)istheLaplacianofGaussianoftheimageItisthec

5、onvolutionoftheGaussiansecondorderderivativewiththeimageLindebergshowedGaussianfunctionisoptimalforscale-spaceanalysisThispaperarguesthatGaussianisoverratedsincethepropertythatnonewstructurescanappearwhilegoingtolowerresolutionisnotprovenin2Dcase6Det

6、ectionApproximatedsecondorderderivativeswithboxfilters(mean/averagefilter)7DetectionScaleanalysiswithconstantimagesize9x9,15x15,21x21,27x2739x39,51x51…1stoctave2ndoctave8DetectionNon-maximumsuppressionandinterpolationBlob-likefeaturedetector9Descrip

7、tionOrientationAssignmentCircularneighborhoodofradius6saroundtheinterestpoint(s=thescaleatwhichthepointwasdetected)Sidelength=4sCost6operationtocomputetheresponsexresponseyresponse10DescriptionDominantorientationTheHaarwaveletresponsesarerepresenteda

8、svectorsSumallresponseswithinaslidingorientationwindowcoveringanangleof60degreeThetwosummedresponseyieldanewvectorThelongestvectoristhedominantorientationSecondlongestis…ignored11DescriptionSplittheinterestregionupinto4x4squaresub-regionswith5x5