版权说明:本文档由用户提供并上传,收益归属内容提供方,若内容存在侵权,请进行举报或认领
文档简介
TowardsRobustObjectDetectionInvarianttoReal-WorldDomainShifts(ICLR2023)
HKUST&MPIIÐZurich
QiFan,MattiaSegu,Yu-WingTai,FisherYu,Chi-KeungTang,
BerntSchiele,DengxinDai
DomainShifts&Generalization
DomainShifts
•Sameclassbutdifferentvisualstyles
SolvingDomainShifts
•DomainAdaptation
•Labeledsourcedomain
•Labeled/unlabeledtargetdomain
•Usebothdomainimagestotrainthemodeltogeneralizeonthetargetdomain
•DomainGeneralization
•Multiple/singlesourcedomain
•Onlyusethesourcedomaintotrainthemodeltogeneralizeontargetdomains
Real-worldDomainShifts
OurTarget
Trainamodel
ononesourcedomain
Applythemodel
ondiversetargetdomains
Real-WorldDomainShifts
ProblemAnalysis
DomainShiftsObservation
•Domainshiftsaremainlyreflectedbystyleshifts.
DomainShiftsObservation
•Domainshiftsaremainlyreflectedbystyleshifts.
•Wecantraindomain-invariantmodelswithdiversesynthesizeddomainstyles.
DomainShiftsObservation
•Domainshiftsaremainlyreflectedbystyleshifts.
•Wecantraindomain-invariantmodelswithdiversesynthesizeddomainstyles.
•Domainstylesareencodedbyfeaturechannelstatistics.
DomainShiftsObservation
•Domainshiftsaremainlyreflectedbystyleshifts.
•Wecantraindomain-invariantmodelswithdiversesynthesizeddomainstyles.
•Domainstylesareencodedbyfeaturechannelstatistics.
•Perturbingfeaturechannelstatisticscansynthesizenewstyles.
(TheResNetbackboneblock1)
PreviousClassificationDG
MethodsFails
ClassificationDGsourcedomain:PACS
DetectionDGsourcedomain:Cityscapes
High
Low
DomainStyleVariance
ImageContextDiversity
Low
High
PreviousClassificationDGMethodsFails
Trainamodel
ononesourcedomain
•Smalldomainstylevariancerestrictsfeature-leveldomainsynthesis.
•Largecontextdiversityrestrictsimage-leveldomainsynthesis.
ImageStyleMatters
x
μ,σ
Replacethefeaturechannelstatistics
μ!ew,σ!ew
ProblemAnalysis
•MixstyleandDSUaresuboptimalforrobustobjectdetectionwhentheinter-imagestylevarianceissmall.
ProblemAnalysis
•MixstyleandDSUaresuboptimalforrobustobjectdetectionwhentheinter-imagestylevarianceissmall.
OurSimpleMethod
NormalizationPerturbation
•AdaptiveInstanceNormalization(AdaIN)
•NormalizationPerturbation
NormalizationPerturbation
.NormalizationPerturbation:y=ax+(β−a)μc
x∈ℛBXCXHXWistheCNNfeatures.
μc∈ℛBXCisthechannelstatistics(mean),estimatedontheinputfeatures.{a,β}∈ℛBXCarerandomnoisesdrawnfromtheGaussiandistribution.
δisthenormalizedstatisticvarianceofthemini-batchofmultiplefeaturechannelstatistics.
NormalizationPerturbation
.NormalizationPerturbation:y=ax+(β−a)μc
xERBXCXHXWistheCNNfeatures.
μcERBXCisthechannelstatistics(mean),estimatedontheinputfeatures.fa,β}ERBXCarerandomnoisesdrawnfromtheGaussiandistribution.
δisthenormalizedstatisticvarianceofthemini-batchofmultiplefeaturechannelstatistics.
21
NormalizationPerturbation
.NormalizationPerturbation:y=ax+(β−a)μc
.NormalizationPerturbationPlus:y=ax+δ(β−a)μc
xERBXCXHXWistheCNNfeatures.
μcERBXCisthechannelstatistics(mean),estimatedontheinputfeatures.fa,β}ERBXCarerandomnoisesdrawnfromtheGaussiandistribution.
δisthenormalizedstatisticvarianceofthemini-batchofmultiplefeaturechannelstatistics.
22
NormalizationPerturbationPlus
•Motivation:somechannelssignificantlyvaryasthedomainchanges
NormalizationPerturbation
Advantages
EffectiveDomainBlending
HighContentFidelity
•Image-leveldomainsynthesis
•maydestroythecontentstructuresoftheoriginalimages
•stylesaredeterministicandlimited
•thestyleaugmentationisonlyperformedonthelow-dimensionalimagespace.
DiverseLatentStyles
BenefitsOtherMethods
AblationStudies
AblationStudies
ComparisonResults
RobustObjectDetection
UnsupervisedDomainAdaptiveObjectDetec
温馨提示
- 1. 本站所有资源如无特殊说明,都需要本地电脑安装OFFICE2007和PDF阅读器。图纸软件为CAD,CAXA,PROE,UG,SolidWorks等.压缩文件请下载最新的WinRAR软件解压。
- 2. 本站的文档不包含任何第三方提供的附件图纸等,如果需要附件,请联系上传者。文件的所有权益归上传用户所有。
- 3. 本站RAR压缩包中若带图纸,网页内容里面会有图纸预览,若没有图纸预览就没有图纸。
- 4. 未经权益所有人同意不得将文件中的内容挪作商业或盈利用途。
- 5. 人人文库网仅提供信息存储空间,仅对用户上传内容的表现方式做保护处理,对用户上传分享的文档内容本身不做任何修改或编辑,并不能对任何下载内容负责。
- 6. 下载文件中如有侵权或不适当内容,请与我们联系,我们立即纠正。
- 7. 本站不保证下载资源的准确性、安全性和完整性, 同时也不承担用户因使用这些下载资源对自己和他人造成任何形式的伤害或损失。
最新文档
- 两家企业合作协议书格式
- 净身出户的离婚协议书应注意啥
- 家庭住宅装潢监理合同范例
- 房屋买卖居间合同书标准格式
- 子女抚养权协议书中的主要内容与要求
- 二手车团购合同范例
- 技术岗位保密协议书模板
- 2024店员工聘用合同范文
- 房地产租赁合同的税务处理
- 山西省协议离婚书范本
- 传染病实验室检查的质量控制
- 期中测试卷(1~3单元)(试题)2024-2025学年五年级上册数学人教版
- 古诗三首《江南春》+公开课一等奖创新教案+教学阐释+素材
- 2024时事政治考试题库(基础题)
- 《学会专注高效学习》初中主题班会课件
- TSDPIA 05-2022 宠物猫砂通用技术规范
- 孙子兵法与兵家智慧
- 果树病虫害防治管理论文
- 油井动液面检测新技术
- 节能工作管理机构和工作职责(经典实用)
- 设备制造流程及制作周期
评论
0/150
提交评论