资源描述:
《Exemplars for Object Detection.pdf》由会员上传分享,免费在线阅读,更多相关内容在学术论文-天天文库。
1、ExemplarsforObjectDetectionNoahSnavelyCS7670:September5,2011Announcements•Officehours:Thursdays1pm–2:30pm•CoursescheduleisnowonlineObjectdetection:wherearewe?Credit:Flickruserneilalderney123•Incredibleprogressinthelasttenyears•Betterfeatures,bettermodels,betterlearningmethods,betterda
2、tasets•CombinationofscienceandhacksThe800-lbGorillaofVisionContests•PASCALVOCChallenge•20categories•Annualclassification,detection,segmentation,…challengesObjectdetectionperformance(2010)Objectdetectionperformance(2010)The2011serveropenedforsubmissionstoday!Machinelearningforobjectdet
3、ection•Whatfeaturesdoweuse?–intensity,color,gradientinformation,…•Whichmachinelearningmethods?–generativevs.discriminative–k-nearestneighbors,boosting,SVMs,…•Whathacksdoweneedtogetthingsworking?HistogramofOrientedGradients(HoG)HoGify10x10cells20x20cells[DalalandTriggs,CVPR2005]Histogr
4、amofOrientedGradients(HoG)HistogramofOrientedGradients(HoG)•LikeSIFT(ScaleInvariantFeatureTransform),but…–Sampledonadense,regulargrid–GradientsarecontrastnormalizedinoverlappingblocksHoGify10x10cells[DalalandTriggs,CVPR2005]20x20cellsHistogramofOrientedGradients(HoG)•Firstusedforappli
5、cationofpersondetection[DalalandTriggs,CVPR2005]•CitedsinceinthousandsofcomputervisionpapersLinearclassifiers•Findlinearfunctiontoseparatepositiveandnegativeexamplesxpositive:xwb0iixnegative:xwb0iiWhichlineisbest?[slidecredit:KristinGrauman]SupportVectorMachines(SVMs)•Discrimina
6、tiveclassifierbasedonoptimalseparatingline(for2Dcase)•Maximizethemarginbetweenthepositiveandnegativetrainingexamples[slidecredit:KristinGrauman]Supportvectormachines•Wantlinethatmaximizesthemargin.xpositive(y1):xwb1iiixnegative(y1):xwb1iiiForsupport,vectors,xwb1iSupportv
7、ectorsMarginC.Burges,ATutorialonSupportVectorMachinesforPatternRecognition,DataMiningandKnowledgeDiscovery,1998[slidecredit:KristinGrauman]Persondetection,ca.20051.Representeachexamplewithasingle,fixedHoGtemplate2.Learnasingle[linear]SVMasadetectorCodeavailable:http://pascal.inrialpes
8、.fr/s