2015_CVPR_Deepid-net Deformable deep convolutional neural networks for object detection

《2015_CVPR_Deepid-net Deformable deep convolutional neural networks for object detection》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、DeepID-Net:DeformableDeepConvolutionalNeuralNetworksforObjectDetectionWanliOuyang,XiaogangWang,XingyuZeng,ShiQiu,PingLuo,YonglongTian,HongshengLi,ShuoYang,ZheWang,Chen-ChangeLoy,XiaoouTangTheChineseUniversityofHongKongwlouyang,xgwang@ee.cuhk.edu.hkAbstractPretrain

2、edonobject-levelannoationPretrainedonimage-levelannotationInthispaper,weproposedeformabledeepconvolutionalImageneuralnetworksforgenericobjectdetection.Thisnewdeeplearningobjectdetectionframeworkhasinnovationsinmultipleaspects.Intheproposednewdeeparchitecture,Visua

3、lizedmodelanewdeformationconstrainedpooling(def-pooling)layermodelsthedeformationofobjectpartswithgeometriccon-(a)straintandpenalty.Anewpre-trainingstrategyisproposedDeepmodelDeformablepatternExamplestolearnfeaturerepresentationsmoresuitablefortheobject...detectio

4、ntaskandwithgoodgeneralizationcapability.Bychangingthenetstructures,trainingstrategies,addingandremovingsomekeycomponentsinthedetectionpipeline,Patternasetofmodelswithlargediversityareobtained,whichDeformationsignificantlyimprovestheeffectivenessofmodelaverag-...pe

5、naltying.TheproposedapproachimprovesthemeanaveragedprecisionobtainedbyRCNN[14],whichwasthestate-of-Patternthe-art,from31%to50.3%ontheILSVRC2014detectiontestset.ItalsooutperformsthewinnerofILSVRC2014,DeformationpenaltyGoogLeNet,by6.1%.Detailedcomponent-wiseanalysis

6、(b)isalsoprovidedthroughextensiveexperimentalevaluation,Figure1.Themotivationofthispaperinnewpretrainingschemewhichprovideaglobalviewforpeopletounderstandthe(a)andjointlylearningfeaturerepresentationanddeformableob-deeplearningobjectdetectionpipeline.jectpartsshar

7、edbymultipleobjectclassesatdifferentsemanticlevels(b).In(a),amodelpretrainedonimage-levelannotationismorerobusttosizeandlocationchangewhileamodelpretrainedonobject-levelannotationisbetterinrepresentingobjectswith1.Introductiontightboundingboxes.In(b),whenipodrotat

8、es,itscircularpatternmoveshorizontallyatthebottomoftheboundingbox.Therefore,Objectdetectionisoneofthefundamentalchallengesinthecircularpatternshavesmall