无需3D数据的开放词汇单目3D物体检测模型训练 Training an Open-Vocabulary Monocular 3D Object Detection Model without 3D Data.docx

下载文档

0
0
约7.19万字
约 34页
2024-12-02 发布于山西
举报
版权申诉
保障服务

无需3D数据的开放词汇单目3D物体检测模型训练 Training an Open-Vocabulary Monocular 3D Object Detection Model without 3D Data.docx

1、本文档共34页，可阅读全部内容。
2、有哪些信誉好的足球投注网站（book118）网站文档一经付费（服务费），不意味着购买了该文档的版权，仅供个人/单位学习、研究之用，不得用于商业用途，未经授权，严禁复制、发行、汇编、翻译或者网络传播等，侵权必究。
3、本站所有内容均由合作方或网友上传，本站不对文档的完整性、权威性及其观点立场正确性做任何保证或承诺！文档内容仅供研究参考，付费前请自行鉴别。如您付费，意味着您自己接受本站规则且自行承担风险，本站不退款、不进行额外附加服务；查看《如何避免下载的几个坑》。如果您已付费下载过本站文档，您可以点击这里二次下载。
4、如文档侵犯商业秘密、侵犯著作权、侵犯人身权等，请点击“版权申诉”（推荐），也可以打举报电话：400-050-0827(电话支持时间：9:00-18:30)。

TraininganOpen-VocabularyMonocular3DObjectDetectionModelwithout3DData

RuiHuang1HenryZheng1YanWang2ZhuofanXia1MarcoPavone2,3GaoHuang1,4*

1DepartmentofAutomation,BNRist,TsinghuaUniversity,China

arXiv:2411.15657v1[cs.CV]23Nov20242NVIDIAResearch,USA3

arXiv:2411.15657v1[cs.CV]23Nov2024

4BeijingAcademyofArtificialIntelligence,China

{hr20,jh-zheng22,xzf23}@mails..cn

yanwan@,pavone@,gaohuang@.cn

https://ovm3d-det.github.io

Abstract

Open-vocabulary3Dobjectdetectionhasrecentlyattractedconsiderableattentionduetoitsbroadapplicationsinautonomousdrivingandrobotics,whichaimstoeffectivelyrecognizenovelclassesinpreviouslyunseendomains.However,existingpointcloud-basedopen-vocabulary3Ddetectionmodelsarelimitedbytheirhighdeploymentcosts.Inthiswork,weproposeanovelopen-vocabularymonocular3Dobjectdetectionframework,dubbedOVM3D-Det,whichtrainsdetectorsusingonlyRGBimages,makingitbothcost-effectiveandscalabletopubliclyavailabledata.Unliketraditionalmethods,OVM3D-Detdoesnotrequirehigh-precisionLiDARor3Dsensordataforeitherinputorgenerating3Dboundingboxes.Instead,itemploysopen-vocabulary2Dmodelsandpseudo-LiDARtoautomaticallylabel3DobjectsinRGBimages,fosteringthelearningofopen-vocabularymonocular3Ddetectors.However,training3Dmodelswithlabelsdirectlyderivedfrompseudo-LiDARisinadequateduetoimpreciseboxesestimatedfromnoisypointcloudsandseverelyoccludedobjects.Toaddresstheseissues,weintroducetwoinnovativedesigns:adaptivepseudo-LiDARerosionandboundingboxrefinementwithpriorknowledgefromlargelanguagemodels.Thesetechniqueseffectivelycalibratethe3DlabelsandenableRGB-onlytrainingfor3Ddetectors.ExtensiveexperimentsdemonstratethesuperiorityofOVM3D-Detover

baselinesinbothindoorandoutdoorscenarios.Thecodewillbe