cvpr18-geometry aware learning of maps for camera用于相机定位几何感知地图学习.pdfVIP

Geometry-AwareLearningofMapsforCameraLocalization

SamarthBrahmbhatt1JinweiGu2KihwanKim2JamesHays1JanKautz2

1{samarth.robo,hays}@2{jinweig,kihwank,jkautz}@

1GeorgiaInstituteofTechnology2NVIDIA

Mapsareakeycomponentinimage-basedcameralo-

calizationandvisualSLAMsystems:theyareusedtoes-

tablishgeometricconstraintsbetweenimages,correctdrift

inrelativeposeestimation,andrelocalizecamerasafterlost

tracking.Theexactdefinitionsofmaps,however,areoften

application-specificandhand-craftedfordifferentscenar-

ios(e.g.3Dlandmarks,lines,planes,bagsofvisualwords).Figure1:Cameralocalizationresultsforoutdoor(left)and

Weproposetorepresentmapsasadeepneuralnetcalledindoor(right)scenesfromtheOxfordRobotCar[38]and7-

MapNet,whichenableslearningadata-drivenmaprep-Scenes[48]datasets.Asshown,priorDNN-basedmethods(e.g.,

resentation.Unlikepriorworkonlearningmaps,Map-PoseNet[32,30,31])resultinnoisyestimations,whiletraditional

Netexploitscheapandubiquitoussensoryinputslikevisualvisualodometrybasedmethods(e.g.,stereoVOorDSO[18])of-

odometryandGPSinadditiontoimagesandfusesthemtendriftovertime.Incontrast,MapNetgivesaccuratecamera

togetherforcameralocalization.Geometricconstraint-poseestimatesbyincludingvariousgeometricconstraintsinDNN

trainingandinference.

pressedbytheseinputs,whichhavetraditionallybeenused

inbundleadjustmentorpose-graphoptimization,arefor-

mulatedaslosstermsinMapN