計算機速識技術 中英互譯

Tenyearsago,computervisionresearchersthoughtthatgettingacomputertotellthedifferencebetweenacatandadogwouldbealmostimpossible,evenwiththesignificantadvanceinthestateofartificialintelligence.Nowwecandoitatalevelgreaterthan99percentaccuracy.Thisiscalledimageclassification—giveitanimage,putalabeltothatimage—andcomputersknowthousandsofothercategoriesaswell.

I’magraduatestudentattheUniversityofWashington,andIworkonaprojectcalledDarknet,whichisaneuralnetworkframeworkfortrainingandtestingcomputervisionmodels.Solet’sjustseewhatDarknetthinksofthisimagethatwehave.Whenwerunourclassifieronthisimage,weseewedon’tjustgetapredictionofdogorcat,weactuallygetspecificbreedpredictions.That’sthelevelofgranularitywehavenow.Andit’scorrect.Mydogisinfactamalamute.

Sowe’vemadeamazingstridesinimageclassification,butwhathappenswhenwerunourclassifieronanimagethatlookslikethis?Well...Weseethattheclassifiercomesbackwithaprettysimilarprediction.Andit’scorrect,thereisamalamuteintheimage,butjustgiventhislabel,wedon’tactuallyknowthatmuchaboutwhat’sgoingonintheimage.Weneedsomethingmorepowerful.Iworkonaproblemcalledobjectdetection,wherewelookatanimageandtrytofindalloftheobjects,putboundingboxesaroundthemandsaywhatthoseobjectsare.Sohere’swhathappenswhenwerunadetectoronthisimage.

Now,withthiskindofresult,wecandoalotmorewithourcomputervisionalgorithms.Weseethatitknowsthatthere’sacatandadog.Itknowstheirrelativelocations,theirsize.Itmayevenknowsomeextrainformation.There’sabooksittinginthebackground.Andifyouwanttobuildasystemontopofcomputervision,sayaself-drivingvehicleoraroboticsystem,thisisthekindofinformationthatyouwant.Youwantsomethingsothatyoucaninteractwiththephysicalworld.Now,whenIstartedworkingonobjectdetection,ittook20secondstoprocessasingleimage.Andtogetafeelforwhyspeedissoimportantinthisdomain,here’sanexampleofanobjectdetectorthattakestwosecondstoprocessanimage.Sothisis10timesfasterthanthe20-seconds-per-imagedetector,andyoucanseethatbythetimeitmakespredictions,theentirestateoftheworldhaschanged,andthiswouldn’tbeveryusefulforanapplication.

Ifwespeedthisupbyanotherfactorof10,thisisadetectorrunningatfiveframespersecond.Thisisalotbetter,butforexample,ifthere’sanysignificantmovement,Iwouldn’twantasystemlikethisdrivingmycar.

Thisisourdetectionsystemrunninginrealtimeonmylaptop.SoitsmoothlytracksmeasImovearoundtheframe,andit’srobusttoawidevarietyofchangesinsize,pose,forward,backward.Thisisgreat.Thisiswhatwereallyneedifwe’regoingtobuildsystemsontopofcomputervision.

Soinjustafewyears,we’vegonefrom20secondsperimageto20millisecondsperimage,athousandtimesfaster.Howdidwegetthere?Well,inthepast,objectdetectionsystemswouldtakeanimagelikethisandsplititintoabunchofregionsandthenrunaclassifieroneachoftheseregions,andhighscoresforthatclassifierwouldbeconsidereddetectionsintheimage.Butthisinvolvedrunningaclassifierthousandsoftimesoveranimage,thousandsofneuralnetworkevaluationstoproducedetection.Instead,wetrainedasinglenetworktodoallofdetectionforus.Itproducesalloftheboundingboxesandclassprobabilitiessimultaneously.Withoursystem,insteadoflookingatanimagethousandsoftimestoproducedetection,youonlylookonce,andthat’swhywecallittheYOLOmethodofobjectdetection.Sowiththisspeed,we’renotjustlimitedtoimages;wecanprocessvideoinrealtime.Andnow,insteadofjustseeingthatcatanddog,wecanseethemmovearoundandinteractwitheachother.

Thisisadetectorthatwetrainedon80differentclassesinMicrosoft’sCOCOdataset.Ithasallsortsofthingslikespoonandfork,bowl,commonobjectslikethat.Ithasavarietyofmoreexoticthings:animals,cars,zebras,giraffes.Andnowwe’regoingtodosomethingfun.We’rejustgoingtogooutintotheaudienceandseewhatkindofthingswecandetect.Doesanyonewantastuffedanimal?Therearesometeddybearsoutthere.Andwecanturndownourthresholdfordetectionalittlebit,sowecanfindmoreofyouguysoutintheaudience.Let’sseeifwecangetthesestopsigns.Wefindsomebackpacks.Let’sjustzoominalittlebit.Andthisisgreat.Andalloftheprocessingishappeninginrealtimeonthelaptop.

Andit’simportanttorememberthatthisisageneralpurposeobjectdetectionsystem,sowecantrainthisforanyimagedomain.Thesamecodethatweusetofindstopsignsorpedestrians,bicyclesinaself-drivingvehicle,canbeusedtofindcancercellsinatissuebiopsy.Andthereareresearchersaroundtheglobealreadyusingthistechnologyforadvancesinthingslikemedicine,robotics.Thismorning,IreadapaperwheretheyweretakingacensusofanimalsinNairobiNationalParkwithYOLOaspartofthisdetectionsystem.Andthat’sbecause

Darknetisopensourceandinthepublicdomain,freeforanyonetouse.

Butwewantedtomakedetectionevenmoreaccessibleandusable,sothroughacombinationofmodeloptimization,networkbinarizationandapproximation,weactuallyhaveobjectdetectionrunningonaphone.

AndI’mreallyexcitedbecausenowwehave