Putting Context into Vision把语境分为视觉课件.ppt

《Putting Context into Vision把语境分为视觉课件.ppt》由会员上传分享，免费在线阅读，更多相关内容在教育资源-天天文库。

1、PuttingContextintoVisionDerekHoiemSeptember15,2019QuestionstoAnswerWhatiscontext?Howiscontextusedinhumanvision?Howiscontextcurrentlyusedincomputervision?ConclusionsWhatiscontext?Anydataormeta-datanotdirectlyproducedbythepresenceofanobjectNearbyimagedataContextWha

2、tiscontext?Anydataormeta-datanotdirectlyproducedbythepresenceofanobjectNearbyimagedataSceneinformationContextContextWhatiscontext?Anydataormeta-datanotdirectlyproducedbythepresenceofanobjectNearbyimagedataSceneinformationPresence,locationsofotherobjectsTreeHowdow

3、eusecontext?AttentionArethereanylivefishinthispicture?CluesforFunctionWhatisthis?CluesforFunctionWhatisthis?Nowcanyoutell?Low-ResScenesWhatisthis?Low-ResScenesWhatisthis?Nowcanyoutell?MoreLow-ResWhataretheseblobs?MoreLow-ResThesamepixels!(acar)Whyiscontextuseful?

4、ObjectsdefinedatleastpartiallybyfunctionTreesgrowingroundBirdscanfly(usually)DoorknobshelpopendoorsWhyiscontextuseful?ObjectsdefinedatleastpartiallybyfunctionContextgivescluesaboutfunctionNotrootedintothegroundnottreeObjectinsky{cloud,bird,UFO,plane,superman}Do

5、orknobsalwaysondoorsWhyiscontextuseful?ObjectsdefinedatleastpartiallybyfunctionContextgivescluesaboutfunctionObjectslikesomescenesbetterthanothersToiletslikebathroomsFishlikewaterWhyiscontextuseful?ObjectsdefinedatleastpartiallybyfunctionContextgivescluesaboutfun

6、ctionObjectslikesomescenesbetterthanothersManyobjectsareusedtogetherand,thus,oftenappeartogetherKettleandstoveKeyboardandmonitorHowiscontextusedincomputervision?Neighbor-basedContextMarkovRandomField(MRF)incorporatescontextualconstraintssites(ornodes)neighborsBlo

7、bsandWords–Carbonetto2019Neighbor-basedcontext(MRF)usefulevenwhentrainingdataisnotfullysupervisedLearnsmodelsofobjectsgivencaptionedimages…DiscriminativeRandomFields–Kumar2019Usingdatasurroundingthelabelsite(notjustatthelabelsite)improvesresultsBuildingsvs.Non-Bu

8、ildingsMulti-scaleConditionalRandomField(mCRF)–He2019Independentdata-basedlabelsRawimageLocalcontextScenecontextmCRFFinaldecisionbasedonClassification(localdat