




已阅读5页,还剩68页未读, 继续免费阅读
版权说明:本文档由用户提供并上传,收益归属内容提供方,若内容存在侵权,请进行举报或认领
文档简介
AdaBoost ItsApplications 主講人 虞台文 Outline OverviewTheAdaBoostAlgorithmHowandwhyAdaBoostworks AdaBoostforFaceDetection AdaBoost ItsApplications Overview Introduction AdaBoost Adaptive Alearningalgorithm Buildingastrongclassifieralotofweakerones Boosting AdaBoostConcept weakclassifiers slightlybetterthanrandom strongclassifier WeakerClassifiers weakclassifiers slightlybetterthanrandom strongclassifier EachweakclassifierlearnsbyconsideringonesimplefeatureTmostbeneficialfeaturesforclassificationshouldbeselectedHowtodefinefeatures selectbeneficialfeatures trainweakclassifiers manage weight trainingsamples associateweighttoeachweakclassifier TheStrongClassifiers weakclassifiers slightlybetterthanrandom strongclassifier Howgoodthestrongonewillbe AdaBoost ItsApplications TheAdaBoostAlgorithm TheAdaBoostAlgorithm Given Initialization For FindclassifierwhichminimizeserrorwrtDt i e Weightclassifier Updatedistribution TheAdaBoostAlgorithm Given Initialization For FindclassifierwhichminimizeserrorwrtDt i e Weightclassifier Updatedistribution Outputfinalclassifier Boostingillustration WeakClassifier1 Boostingillustration WeightsIncreased Boostingillustration WeakClassifier2 Boostingillustration WeightsIncreased Boostingillustration WeakClassifier3 Boostingillustration Finalclassifierisacombinationofweakclassifiers AdaBoost ItsApplications HowandwhyAdaBoostworks TheAdaBoostAlgorithm Given Initialization For FindclassifierwhichminimizeserrorwrtDt i e Weightclassifier Updatedistribution Outputfinalclassifier WhatgoaltheAdaBoostwantstoreach TheAdaBoostAlgorithm Given Initialization For FindclassifierwhichminimizeserrorwrtDt i e Weightclassifier Updatedistribution Outputfinalclassifier WhatgoaltheAdaBoostwantstoreach Goal Minimizeexponentialloss Finalclassifier Goal Minimizeexponentialloss Finalclassifier MaximizethemarginyH x Goal Finalclassifier Minimize Define with Then Finalclassifier Minimize Define with Then Set 0 Finalclassifier Minimize Define with Then 0 with Finalclassifier Minimize Define Then 0 with Finalclassifier Minimize Define Then 0 with Finalclassifier Minimize Define Then with Finalclassifier Minimize Define Then with Finalclassifier Minimize Define Then maximizedwhen with Finalclassifier Minimize Define Then Attimet with Finalclassifier Minimize Define Then Attimet Attime1 Attimet 1 AdaBoost ItsApplications AdaBoostforFaceDetection TheTaskofFaceDetection ManyslidesadaptedfromP Viola BasicIdea Slideawindowacrossimageandevaluateafacemodelateverylocation Challenges Slideawindowacrossimageandevaluateafacemodelateverylocation Slidingwindowdetectormustevaluatetensofthousandsoflocation scalecombinations Facesarerare 0 10perimageForcomputationalefficiency weshouldtrytospendaslittletimeaspossibleonthenon facewindowsAmegapixelimagehas 106pixelsandacomparablenumberofcandidatefacelocationsToavoidhavingafalsepositiveineveryimageimage ourfalsepositiveratehastobelessthan10 6 TheViola JonesFaceDetector Aseminalapproachtoreal timeobjectdetectionTrainingisslow butdetectionisveryfastKeyideasIntegralimagesforfastfeatureevaluationBoostingforfeatureselectionAttentionalcascadeforfastrejectionofnon facewindows P ViolaandM Jones Rapidobjectdetectionusingaboostedcascadeofsimplefeatures CVPR2001 P ViolaandM Jones Robustreal timefacedetection IJCV57 2 2004 ImageFeatures Rectanglefilters ImageFeatures Rectanglefilters SizeofFeatureSpace Howmanynumberofpossiblerectanglefeaturesfora24x24detectionregion Rectanglefilters A B C D FeatureSelection Howmanynumberofpossiblerectanglefeaturesfora24x24detectionregion A B C D Whatfeaturesaregoodforfacedetection FeatureSelection Howmanynumberofpossiblerectanglefeaturesfora24x24detectionregion A B C D Canwecreateagoodclassifierusingjustasmallsubsetofallpossiblefeatures Howtoselectsuchasubset Integralimages Theintegralimagecomputesavalueateachpixel x y thatisthesumofthepixelvaluesaboveandtotheleftof x y inclusive x y ComputingtheIntegralImage Theintegralimagecomputesavalueateachpixel x y thatisthesumofthepixelvaluesaboveandtotheleftof x y inclusive Thiscanquicklybecomputedinonepassthroughtheimage x y ComputingSumwithinaRectangle D B C A Only3additionsarerequiredforanysizeofrectangle Scaling Integralimageenablesustoevaluateallrectanglesizesinconstanttime Therefore noimagescalingisnecessary Scaletherectangularfeaturesinstead 1 2 3 4 5 6 Boosting BoostingisaclassificationschemethatworksbycombiningweaklearnersintoamoreaccurateensembleclassifierAweaklearnerneedonlydobetterthanchanceTrainingconsistsofmultipleboostingroundsDuringeachboostinground weselectaweaklearnerthatdoeswellonexamplesthatwerehardforthepreviousweaklearners Hardness iscapturedbyweightsattachedtotrainingexamples Y FreundandR Schapire Ashortintroductiontoboosting JournalofJapaneseSocietyforArtificialIntelligence 14 5 771 780 September 1999 TheAdaBoostAlgorithm Given Initialization For FindclassifierwhichminimizeserrorwrtDt i e Weightclassifier Updatedistribution TheAdaBoostAlgorithm Given Initialization For FindclassifierwhichminimizeserrorwrtDt i e Weightclassifier Updatedistribution Outputfinalclassifier WeakLearnersforFaceDetection Given Initialization For FindclassifierwhichminimizeserrorwrtDt i e Weightclassifier Updatedistribution Outputfinalclassifier Whatbaselearnerisproperforfacedetection WeakLearnersforFaceDetection Boosting TrainingsetcontainsfaceandnonfaceexamplesInitially withequalweightForeachroundofboosting EvaluateeachrectanglefilteroneachexampleSelectbestthresholdforeachfilterSelectbestfilter thresholdcombinationReweightexamplesComputationalcomplexityoflearning O MNK Mrounds Nexamples Kfeatures FeaturesSelectedbyBoosting Firsttwofeaturesselectedbyboosting Thisfeaturecombinationcanyield100 detectionrateand50 falsepositiverate ROCCurvefor200 FeatureClassifier A200 featureclassifiercanyield95 detectionrateandafalsepositiverateof1in14084 Notgoodenough Tobepracticalforrealapplication thefalsepositiveratemustbecloserto1in1 000 000 AttentionalCascade Classifier3 Classifier2 Classifier1 IMAGESUB WINDOW T T Westartwithsimpleclassifierswhichrejectmanyofthenegativesub windowswhiledetectingalmostallpositivesub windowsPositiveresponsefromthefirstclassifiertriggerstheevaluationofasecond morecomplex classifier andsoonAnegativeoutcomeatanypointleadstotheimmediaterejectionofthesub window AttentionalCascade Classifier3 Classifier2 Classifier1 IMAGESUB WINDOW T T Chainclassifiersthatareprogressivelymorecomplexandhavelowerfalsepositiverates DetectionRateandFalsePositiveRateforChainedClassifiers Classifier3 Classifier2 Classifier1 IMAGESUB WINDOW T T ThedetectionrateandthefalsepositiverateofthecascadearefoundbymultiplyingtherespectiveratesoftheindividualstagesAdetectionrateof0 9andafalsepositiverateontheorderof10 6canbeachievedbya10 stagecascadeifeachstagehasadetectionrateof0 99 0 9910 0 9 andafalsepositiverateofabout0 30 0 310 6 10 6 TrainingtheCascade SettargetdetectionandfalsepositiveratesforeachstageKeepaddingfeaturestothecurrentstageuntilitstargetrateshavebeenmetNeedtolowerAdaBoostthresholdtomaximizedetection asopposedtominimizingtotalclassificationerror TestonavalidationsetIftheoverallfalsepositiverateisnotlowenough thenaddanotherstageUsefalsepositivesfromcurrentstageasthenegativetrainingexamplesforthenextstage TrainingtheCascade ROCCurvesCascadedClassifiertoMonlithicClassifier ROCCurvesCascadedClassifiertoMonlithicClassifier Thereislittledifferencebetweenthetwointermsofaccuracy Thereisabigdifferenceintermsofspeed Thecascadedclassifierisnearly10timesfastersinceitsfirststagethrowsoutmostnon facessothattheyareneverevaluatedbysubsequentstages TheImplementedSystem TrainingData5000facesAllfrontal rescaledto24x24pixels300millionnon faces9500non faceimagesFacesarenormalizedScale translationManyvariationsAcrossindividualsIlluminationPose StructureoftheDetectorCascade Combiningsuccessivelymorecomplexclassifiersincascade38stagesincludedatotalof6060features StructureoftheDetectorCascade RejectSub Window 1 2 3 4 5 6 7 8 38 Face F F F F F F F F F T T T T T T T T T AllSub Windows 2features reject50 non faces detect100 faces 10features reject80 non faces detect100 faces 25features 50features byalgorithm SpeedoftheFinalDetector Ona700MhzPentiumIIIprocessor thefacedetectorcanprocessa384 288pixelimageinabout 067seconds 15Hz15timesfasterthanpreviousdetectorofcomparableaccuracy Rowleyetal 1998 Averageof8fea
温馨提示
- 1. 本站所有资源如无特殊说明,都需要本地电脑安装OFFICE2007和PDF阅读器。图纸软件为CAD,CAXA,PROE,UG,SolidWorks等.压缩文件请下载最新的WinRAR软件解压。
- 2. 本站的文档不包含任何第三方提供的附件图纸等,如果需要附件,请联系上传者。文件的所有权益归上传用户所有。
- 3. 本站RAR压缩包中若带图纸,网页内容里面会有图纸预览,若没有图纸预览就没有图纸。
- 4. 未经权益所有人同意不得将文件中的内容挪作商业或盈利用途。
- 5. 人人文库网仅提供信息存储空间,仅对用户上传内容的表现方式做保护处理,对用户上传分享的文档内容本身不做任何修改或编辑,并不能对任何下载内容负责。
- 6. 下载文件中如有侵权或不适当内容,请与我们联系,我们立即纠正。
- 7. 本站不保证下载资源的准确性、安全性和完整性, 同时也不承担用户因使用这些下载资源对自己和他人造成任何形式的伤害或损失。
最新文档
- 璀璨的香港课件
- 下交叉综合征的治疗方案
- 培训学校的运营
- 2025年麻醉学围手术期镇痛管理挑战考核答案及解析
- 辅警基本业务知识培训课件
- 2025年3D打印技术的快速成型工艺优化
- 2025年3D打印的快速制造技术
- 交通银行2025来宾市秋招笔试EPI能力测试题专练及答案
- 农业银行2025庆阳市秋招笔试性格测试题专练及答案
- 工商银行2025韶关市秋招笔试性格测试题专练及答案
- 第2课 第二次鸦片战争 课件 部编版2024八年级上册历史
- 艾梅乙信息安全管理制度
- 心包疾病患者的护理
- 高校援建人才管理办法
- 企业创新激励机制与绩效评价体系研究
- 《老年人能力评估实务》健康养老服务专业全套教学课件
- 橡胶环境损伤机理与寿命预测模型构建研究
- 腹腔镜子宫肌瘤护理查房
- 2025年卫生类事业单位招聘考试护理学专业知识外科护理试卷
- 个人养老金微课课件
- 肿瘤患者心理抑郁护理
评论
0/150
提交评论