资源描述：
《SemanticFusion Dense 3D Semantic Mapping with Convolutional Neural Networks》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、SemanticFusion:Dense3DSemanticMappingwithConvolutionalNeuralNetworksJohnMcCormac,AnkurHanda,AndrewDavison,andStefanLeuteneggerDysonRoboticsLab,ImperialCollegeLondonAbstract—Evermorerobust,accurateanddetailedmappingusingvisualsensinghasproventobeanenablingfactorformobilerobotsacros
2、sawidevarietyofapplications.Forthenextlevelofrobotintelligenceandintuitiveuserinteraction,mapsneedextendbeyondgeometryandappearence—theyneedtocontainsemantics.WeaddressthischallengebycombiningConvolutionalNeuralNetworks(CNNs)andastateoftheartdenseSimultaneousLocalisationandMapping
3、(SLAM)system,ElasticFusion,whichprovideslong-termdensecorrespondencebetweenframesofindoorRGB-Dvideoevenduringloopyscanningtrajectories.ThesecorrespondencesallowtheCNN’sFig.1:Theoutputofoursystem:Ontheleft,adensesurfelsemanticpredictionsfrommultipleviewpointstobeproba-basedreconstr
4、uctionfromavideosequenceintheNYUv2bilisticallyfusedintoamap.Thisnotonlyproducesausefultestset.Ontherightthesamemap,semanticallyannotatedsemantic3Dmap,butwealsoshowontheNYUv2datasetthatfusingmultiplepredictionsleadstoanimprovementeveninthewiththeclassesgiveninthelegendbelow.2Dseman
5、ticlabellingoverbaselinesingleframepredictions.Wealsoshowthatforasmallerreconstructiondatasetwithlargervariationinpredictionviewpoint,theimprovementoversingleframesegmentationincreases.Oursystemisefficientenoughsurfelsremainpersistentlyassociatedwithreal-worldentitiestoallowreal-ti
6、meinteractiveuseatframe-ratesof25Hz.andthisenableslong-termfusionofper-framesemanticpredictionsoverwidechangesinviewpoint.I.INTRODUCTIONThegeometryofthemapitselfalsoprovidesusefulTheinclusionofrichsemanticinformationwithinadenseinformationwhichcanbeusedtoefficientlyregularisethema
7、penablesamuchgreaterrangeoffunctionalitythanfinalpredictions.Ourpipelineisdesignedtoworkonline,andgeometryalone.Forinstance,indomesticrobotics,asimplealthoughwehavenotfocusedonperformance,theefficiencyfetchingtaskrequiresknowledgeofbothwhatsomethingis,ofeachcomponentleadstoareal-tim
8、ecapable(25Hz)aswellaswhereitisl
