资源描述：
《Fully-Convolutional Point 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、Fully-ConvolutionalPointNetworksforLarge-ScalePointCloudsDarioRethage1,JohannaWald1,J•urgenSturm2,NassirNavab1,andFedericoTombari11TechnicalUniversityofMunichdario.rethage@tum.de,johanna.wald@tum.de,navab@cs.tum.de,tombari@in.tum.de2Googlejsturm@google.comAbstract.Thisworkproposesag
2、eneral-purpose,fully-convolutionalnetworkarchitectureforecientlyprocessinglarge-scale3Ddata.Onestrikingcharacteristicofourapproachisitsabilitytoprocessunorga-nized3Drepresentationssuchaspointcloudsasinput,thentransform-ingtheminternallytoorderedstructurestobeprocessedvia3Dcon-volut
3、ions.Incontrasttoconventionalapproachesthatmaintaineitherunorganizedororganizedrepresentations,frominputtooutput,ourap-proachhastheadvantageofoperatingonmemoryecientinputdatarepresentationswhileatthesametimeexploitingthenaturalstructureofconvolutionaloperationstoavoidtheredundantco
4、mputingandstoringofspatialinformationinthenetwork.Thenetworkeliminatestheneedtopre-orpostprocesstherawsensordata.This,togetherwiththefully-convolutionalnatureofthenetwork,makesitanend-to-endmethodabletoprocesspointcloudsofhugespacesorevenentireroomswithupto200kpointsatonce.Anotherad
5、vantageisthatournetworkcanpro-duceeitheranorderedoutputormappredictionsdirectlyontotheinputcloud,thusmakingitsuitableasageneral-purposepointclouddescriptorapplicabletomany3Dtasks.Wedemonstrateournetwork'sabilitytoe ectivelylearnbothlow-levelfeaturesaswellascomplexcompositionalrelati
6、onshipsbyevaluatingitonbenchmarkdatasetsforsemanticvoxelsegmentation,semanticpartsegmentationand3Dscenecaptioning.Keywords:PointClouds3DDeepLearningSceneUnderstandingFully-ConvolutionalSemanticSegmentation3DCaptioningarXiv:1808.06840v1[cs.CV]21Aug20181IntroductionProcessing3Dda
7、taasobtainedfrom3Dscannersordepthcamerasisfun-damentaltoawealthofapplicationsinthe eldof3Dcomputervision,sceneunderstanding,augmented/mixedreality,roboticsandautonomousdriving.Theextractionofreliablesemanticinformationfroma3Dsceneisuseful,forexam-ple,toappropriatelyaddvirtualcontent
8、tothe3Dspacearoundu
