Bayesian Information Recovery from CNN for Probabilistic Inference

Bayesian Information Recovery from CNN for Probabilistic Inference

ID：40363211

大?。?.24 MB

頁數(shù)：8頁

時間：2019-08-01

Bayesian Information Recovery from CNN for Probabilistic Inference_第1頁
Bayesian Information Recovery from CNN for Probabilistic Inference_第2頁
Bayesian Information Recovery from CNN for Probabilistic Inference_第3頁
Bayesian Information Recovery from CNN for Probabilistic Inference_第4頁
Bayesian Information Recovery from CNN for Probabilistic Inference_第5頁
資源描述：

《Bayesian Information Recovery from CNN for Probabilistic Inference》由會員上傳分享，免費在線閱讀，更多相關(guān)內(nèi)容在學(xué)術(shù)論文-天天文庫。

1、2018IEEE/RSJInternationalConferenceonIntelligentRobotsandSystems(IROS)Madrid,Spain,October1-5,2018BayesianInformationRecoveryfromCNNforProbabilisticInferenceDmitryKopitkovandVadimIndelmanAbstract—Typicalinferenceapproachesthatworkwithhigh-dimensionalvisu

2、almeasurementsusehand-engineeredimagefeatures(e.g.SIFT)thatrequirecombinatorialdataassociation,orpredictonlyhiddenstatemeanwithoutconsideringitsuncertaintyandmulti-modalityaspects.WedevelopanovelapproachtoinfersystemhiddenstatefromvisualobservationsviaCN

3、NfeatureswhichareoutputsofaCNNclassi?er.Tothatend,atpre-deploymentstageweuseneuralnetworksto(a)learnagenerativeviewpoint-dependentmodelofCNNfeaturesgiventherobotposeandapproximatethismodelbyaspatially-varyingGaussiandistribution.Further,atdeploymentthism

4、odelisutilizedwithinaBayesianframeworkforproba-bilisticinference,consideringarobotlocalizationproblem.Ourmethoddoesnotinvolvedataassociationandprovidesuncertaintycovarianceofthe?nalestimation.Moreover,weshowempiricallythattheCNNfeaturelikelihoodisunimoda

5、lwhichsimpli?estheinferencetask.WetestourmethodinasimulatedUnrealEngineenvironment,wherewesucceedtoretrievehigh-levelstateinformationfromCNNfeaturesandproducetrajectoryestimationwithhighaccuracy.Additionally,weanalyzerobustnessofourapproachtodifferentlig

6、htconditions.I.INTRODUCTIONInferringasystemstatefrommultiplemeasurements,pos-siblycapturedbydifferentsensors,isafundamentalproblem(b)Fig.1:Approachoverview.InthispaperweuseCNNfeaturesforrobot’sstateinrobotics.Bayesianinferenceforsystemidenti?cationisinfe

7、rencewithinaBayesianframework.Animagecapturedfromrobotposexioneofthemainbuildingblocksonwhichmodernreal-ispassedtoaCNNclassi?erwhichproducesafeaturesvectorfithatrepresentstheimage.(a)Duringthepre-deploymentstagewelearnspatially-varyingCNNworldroboticappl

8、icationsrely,suchasautonomousnavi-probabilitylikelihoodP(fijxi)approximatedbyN((xi);(xi)).Twoneuralgationandsimultaneouslocalizationandmapping(SLAM).networksproduceviewpoint-dependentmeanandcovariancefunctionsoffigivenxi

當前文檔最多預(yù)覽五頁，下載文檔查看全文

此文檔下載收益歸作者所有

當前文檔最多預(yù)覽五頁，下載文檔查看全文
溫馨提示：
1. 部分包含數(shù)學(xué)公式或PPT動畫的文件，查看預(yù)覽時可能會顯示錯亂或異常，文件下載后無此問題，請放心下載。
2. 本文檔由用戶上傳，版權(quán)歸屬用戶，天天文庫負責(zé)整理代發(fā)布。如果您對本文檔版權(quán)有爭議請及時聯(lián)系客服。
3. 下載前請仔細閱讀文檔內(nèi)容，確認文檔內(nèi)容符合您的需求后進行下載，若出現(xiàn)內(nèi)容與標題不符可向本站投訴處理。
4. 下載文檔時可能由于網(wǎng)絡(luò)波動等原因無法下載或下載錯誤，付費完成后未能成功下載的用戶請聯(lián)系客服處理。