研究boosting必看
In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Minneapolis, MN, 2007
OnlineLearningAsymmetricBoostedClassi ersforObjectDetection
Minh-TriPham
Tat-JenCham
SchoolofComputerEngineeringNanyangTechnologicalUniversity
Singapore{mtpham,astjcham}@ntu.edu.sg
Abstract
Wepresentanintegratedframeworkforlearningasym-metricboostedclassi ersandonlinelearningtoaddresstheproblemofonlinelearningasymmetricboostedclassi ers,whichisapplicabletoobjectdetectionproblems.Inpar-ticular,ourmethodseekstobalancetheskewnessofthelabelspresentedtotheweakclassi ers,allowingthemtobetrainedmoreequally.Inonlinelearning,weintroduceanextraconstraintwhenpropagatingtheweightsofthedatapointsfromoneweakclassifertoanother,allowingthealgorithmtoconvergefaster.IncomparedwiththeOn-lineBoostingalgorithmrecentlyappliedtoobjectdetectionproblems,weobservedabout0-10%increaseinaccuracy,andabout5-30%gaininlearningspeed.
However,inthetasksofobjectdetectionandrecogni-tion(e.g.,facedetectionordatabaseretrieval),oneoftenfaceswithabinaryclassi cationproblemwheretheproba-bilityofobservingthepositiveexamples(i.e.thetargetob-ject)ismuchlowerthantheprobabilityofobservingthenegativeexamples(e.g.background,non-targetobject).Insuchdomains,anasymmetricclassi erthatcanavoidthedangerofmissingpositivepatternsisoftenmoredesirablethantheoneminimizingtheclassi cationerror.Andthen,highdetectionratescanbeachievedbycascadingtheclas-si ers(e.g.[15,16])orhierarchicallyorganizingthem(e.g.[8,10])fromcoarseto ne.
Therehavebeenworksaddressingtheproblemoflearn-ingasymmetricclassi ers[1,7,11,16].Thoughbeingdif-feredindetails,thecommonapproachofthesemethodsistosetupanasymmetricexpectedlosswherefalsenegativesarepenalizedmorethanfalsepositives.ViolaandJones[16]introducedanasymmetricloss:falsenegativescostsktimesmorethanfalsepositives.Theyappliedanasym-metricre-weightingtechniquebeforetrainingeachweakclassi ersothattheresultisequivalenttominimizingtheirasymmetricloss.MaandDing[11]appliedcost-sensitivelearningtechnique[1]tofacedetectionwithre-weightingschemesimilartothatofViolaandJones[16].However,theweight-updatingrule[1]involvesaparametercwhichisapplication-dependentandneedsextensivetrialsforbestperformance.Houetal.[7]proposedaninterestingadap-tivetechniqueavoidingtheneedforapplication-dependentparameters,buttheirmethodrequiresre-trainingtheweakclassi ersmultipletimesto ne-tunetheirasymmetricpa-rameterat.Toourbestknowledge,theirmethodcouldnotbedoneonline.
Inthispaper,weintroduceanovelboostingalgorithmwhichaddressestheproblemoflearningonlineanasym-metricboostedclassi er.Theideaoflearningonlineissim-ilartotheOnlineBoostingalgorithm[12]havingbeenbeenrecentlyused[5,9,13],butwithafasterconvergencespeed.Ontheaspectoflearningasymmetrically,ourmethodis
1.Introduction
AdaBoost[3,14]hasbeensuccessfullyusedinmanymachinelearningandpatternrecognitiontasksincom-putervision.Itsunderlyingideaistocombineanensem-bleofMweaklearnerstoproducethe nalclassi erwithveryhighaccuracy.Inthetasksofobjectdetectionandrecognition,especiallyforfacedetection,impressiveresults[8,10,15,16]havebeenreportedwhencascadingorhier-archicallyorganizingsuchboostedclassi ersfromcoarseto ne,achievinga nalclassi erwithhighdetectionratesandfastdetectionspeed.
Recently,therehasbeenconsiderableinterestinapply-ingboostingtechniquesoncomputervisionproblemsthatrequireonlinelearning.Examplesare:onlineselectionofdiscriminativefeaturesofGrabnerandBischof[5],onlineconservativelearningofRothetal.[13],andonlineco-trainingofJavedetal.[9].Thesemethodsusethesameun-derlyingOnlineBoostingalgorithmproposedbyOza[12]tolearnonline.Thealgorithmminimizestheclassi cationerrorwhileupdatingtheweakclassi ersonline.
提出一种新的策略在 false positive 和检测率中取得平衡 Learning Boosted Asymmetric Classifiers for Object Detection Proceedings of the 2006 IEEE Computer Society...
Rapid object detection u... 8页 免费 Online Learning Asymmetr... 8页 ... Empirical Analysis of Detection Cascades of Boosted Classifiers for Rapid ...
? A machine learning approach for visual object detection ? ? Capable of ...A combining classifiers method ? cascade classifiers Discard the background ...
SCHAPIRE 和 YORAM SINGER,于 Machine Learning 发表论文: Improved Boosting ...Detection Cascades of Boosted Classifiers for Rapid Object Detection 中提出在 ...
Rapid Object Detection using a Boosted Cascade of...? ? A machine learning algorithm Perform ...Classifiers Weight (α) ffinal(x) = α final,...
Modeling, Learning, Computing, and Sampling, Vancouver, Canada, July 13, ... of detection cascades of boosted classifiers for rapid object detection. ...
Modeling, Learning, Computing, and Sampling, Vancouver, Canada, July 13, ... of detection cascades of boosted classifiers for rapid object detection. ...
too. See "Empirical Analysis of Detection Cascades of Boosted Classifiers for Rapid Object Detection" technical report by R.Lienhart et al for empirical ...
SCHAPIRE 和 YORAM SINGER,于 Machine Learning 发表论文: Improved Boosting ... of Detection Cascades of Boosted Classifiers for Rapid Object Detection. ...
too. See “Empirical Analysis of Detection Cascades of Boosted Classifiers for Rapid Object Detection” technical report by R.Lienhart et al for empirical ...
热门文档