Convolutional Neural Network Based Monocular Depth Estimation

Abstract

Foriailongitime,istereoicamerasihaveibeenideployediinivisualisimultaneousi locationiandimappingi(SLAM)isystemsitoiobtaini3Diinformation.iAlthoughistereoi camerasishowigoodiperformance,itheimainidrawbackiisitheicomplexiandi expensiveihardwareisetupiitirequires,iwhichirestrictsitheiuseiofitheisystem.i Monocularicamerasiareia simpleriandicheaperialternative.iRecentiworkihasi shown ithatiaccessitoidepthimapsiinitheimonocularisystemiisibeneficialiasitheyicanibeiuse ditoiimprovei3Direconstruction.iThisiworkiproposesiaideepineuralinetworkithatipr edictsidenseihigh-resolutionidepthimapsifromimonoculariRGBiimages.i Networkiarchitectureifollowsianiencoder-decoderistructureiiniwhichimultiscaleiinformationiisicapturediandiskip-connectionsiareiuseditoiretrieveidetails.i TheinetworkiisitrainediandievaluatedioniaiNYUiv2idatasetiwithiresultsi comparableitoistate-ofi-ithe-artimethods.iTheiproblemiofidepthiestimationiisiani importanticomponentiforiunderstandingitheigeometryiofiaisceneiandifori navigatingithroughispace.iMoreiunderstandingiofitheienvironment,isuchi asirecognitioniactivities,Icontributesitoichangesiiniotherifields. Inimanyiapplications,iaccurateimeasurementiofidepthifromiimagesiisiaicrucialitask involvingiinterpretationiandirestorationiofitheiscene.iExistingimethodsifori calculatingidepthiofteniyieldifuzzyiapproximationsiwithilowiresolution.iThisi thesisidescribesiaiconvolutionineuralinetworkitoiuseitransferilearningitoi computeiaihighiresolutionimapiofidepthiwithioneisingleiRGBiimage.iUsingiai typicaliencoder-decoderimodel,iwheniinitializingiouriencoder,iweiexploitifeaturesi derivediusingihighiperformanceipre-trainedinetworksialongiwithiextensioniandi testing techniques ithatiresultiinimoreiaccurateiresults. xiii Weidemonstrateihowiouriapproachicaniachieveiaccurate,ihigh-resolutioni depthimaps,ieveniforiaiveryisimpleidecoder.iWeitrainidatasetionithreeimodelsii.e.i denseneti169,i201iandiResnet50.iOurinetworkiconductsiresultsiwithistateiofitheiart ionitwoidatasetsiwithilessiparametersianditrainingiiterations,iandialsoioffersi qualitativelyibetteritestsithatireflectihumaniboundariesimoreiaccurately.iOuri algorithmigivesistateiofitheiartiresultsiandiitigivesirmsivalueiofi0.4611.