资源描述：
《Single Image Depth Estimation》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、SingleImageDepthEstimationFromPredictedSemanticLabelsBeyangLiuStephenGouldDaphneKollerDept.ofComputerScienceDept.ofElectricalEngineeringDept.ofComputerScienceStanfordUniversityStanfordUniversityStanfordUniversitybeyangl@cs.stanford.edusgould@stanford.edukoll
2、er@cs.stanford.eduAbstractWeconsidertheproblemofestimatingthedepthofeachpixelinascenefromasinglemonocularimage.Unliketra-ditionalapproaches[18,19],whichattempttomapfromappearancefeaturestodepthdirectly,wefirstperformasemanticsegmentationofthesceneandusethesem
3、anticlabelstoguidethe3Dreconstruction.Thisapproachpro-videsseveraladvantages:ByknowingthesemanticclassFigure1.Exampleoutputfromourmodelshowinghowsemanticofapixelorregion,depthandgeometryconstraintscanclassprediction(center)stronglyinformsdepthperception(righ
4、t).beeasilyenforced(e.g.,“sky”isfarawayand“ground”Semanticclassesareshownoverlayedonimage.Depthindicatedishorizontal).Inaddition,depthcanbemorereadilypre-bycolormap(redismoredistant).SeeFigure6forcolorlegend.dictedbymeasuringthedifferenceinappearancewithre-s
5、pecttoagivensemanticclass.Forexample,atreewillhavemoreuniformappearanceinthedistancethanitdoesmated3Dscenereconstruction[19,12,4,11,18]hasfo-closeup.Finally,theincorporationofsemanticfeaturescusesonextractingthesegeometriccuesandadditionalin-allowsustoachiev
6、estate-of-the-artresultswithasignifi-formationfromnovelimages.cantlysimplermodelthanpreviousworks.Theseworkslargelyignorethetaskofsemanticunder-standingandjumpstraighttoestimatingdepthorgeometryfromimagefeatures.Usingmachinelearningtechniques,1.Introductionth
7、eseapproachesdetermineadirectmappingfromimageRecoveringthe3Dstructureofascenefromasinglefeaturestodepth.However,thisputsanenormousburdenimageisafundamentalproblemincomputervisionthatonthelearningalgorithm,whichmustnowimplicitlyrea-hasapplicationinrobotics,su
8、rveillanceandgeneralscenesonaboutsemanticcontext(e.g.,thedifferencebetweenvi-understanding—ifwecanestimatescenestructurethenwesuallysimilarpatchesofskyandwater)toaccuratelylearncanbetterundersta
