图像处理_INRIA Person Dataset(法国国家信息与自动化研究所行人数据

INRIAPersonDataset(法国国家信息与自动化研究所行人数据库)数据摘要：Thisdatasetwascollectedaspartofresearchworkondetectionofuprightpeopleinimagesandvideo.TheresearchisdescribedindetailinCVPR2005paperHistogramsofOrientedGradientsforHumanDetectionandmyPhDthesis.Thedatasetisdividedintwoformats:(a)originalimageswithcorrespondingannotationfiles,and(b)positiveimagesinnormalized64x128pixelformat(asusedintheCVPRpaper)withoriginalnegativeimages.Thedatasetcontainsimagesfromseveraldifferentsources:ImagesfromGRAZ01dataset,thoughannotationfilesarecompletelynew.Imagesfrompersonaldigitalimagecollectionstakenoveralongtimeperiod.Usuallytheoriginalpositiveimageswereofveryhighresolution(approx.2592x1944pixels),sowehavecroppedtheseimagestohighlightpersons.Manypeoplearebystanderstakenfromthebackgroundsoftheseinputphotos,soideallythereisnoparticularbiasintheirpose.Fewofimagesaretakenfromthewebusinggoogleimages.中文关键词：直立,行人,检测,标注,规范化,英文关键词：upright,people,detection,annotation,normalized,数据格式：IMAGE数据用途：Researchworkondetectionofuprightpeopleinimagesandvideo数据详细介绍：INRIAPersonDatasetThisdatasetwascollectedaspartofresearchworkondetectionofuprightpeopleinimagesandvideo.TheresearchisdescribedindetailinCVPR2005paperHistogramsofOrientedGradientsforHumanDetectionandmyPhDthesis.Thedatasetisdividedintwoformats:(a)originalimageswithcorrespondingannotationfiles,and(b)positiveimagesinnormalized64x128pixelformat(asusedintheCVPRpaper)withoriginalnegativeimages.ContributionsThedatasetcontainsimagesfromseveraldifferentsources:ImagesfromGRAZ01dataset,thoughannotationfilesarecompletelynew.Imagesfrompersonaldigitalimagecollectionstakenoveralongtimeperiod.Usuallytheoriginalpositiveimageswereofveryhighresolution(approx.2592x1944pixels),sowehavecroppedtheseimagestohighlightpersons.Manypeoplearebystanderstakenfromthebackgroundsoftheseinputphotos,soideallythereisnoparticularbiasintheirpose.Fewofimagesaretakenfromthewebusinggoogleimages.NoteOnlyuprightpersons(withpersonheight100)aremarkedineachimage.Annotationsmaynotberight;inparticularattimesportionsofannotatedboundingboxesmaybeoutsideorinsidetheobject.OriginalImagesFolders'Train'and'Test'correspond,respectively,tooriginaltrainingandtestimages.Bothfoldershavethreesubfolders:(a)'pos'(positivetrainingortestimages),(b)'neg'(negativetrainingortestimages),and(c)'annotations'(annotationfilesforpositiveimagesinPascalChallengeformat).NormalizedImagesFolders'train_64x128_H96'and'test_64x128_H96'correspondtonormalizeddatasetasusedinabovereferencedpaper.Bothfoldershavetwosubfolders:(a)'pos'(normalizedpositivetrainingortestimagescenteredonthepersonwiththeirleft-rightreflections),(b)'neg'(containingoriginalnegativetrainingortestimages).Noteimagesinfolder'train/pos'areof96x160pixels(amarginof16pixelsaroundeachside),andimagesinfolder'test/pos'areof70x134pixels(amarginof3pixelsaroundeachside).Thishasbeendonetoavoidboundaryconditions(thustoavoidanyparticularbiasintheclassifier).Inbothfolders,usethecentered64x128pixelswindowfororiginaldetectiontask.NegativewindowsTogeneratenegativetrainingwindowsfromnormalizedimages,afixedsetof12180windows(10windowspernegativeimage)aresampledrandomlyfrom1218negativetrainingphotosprovidingtheinitialnegativetrainingset.Foreachdetectorandparametercombination,apreliminarydetectoristrainedandallnegativetrainingimagesaresearchedexhaustively(overascale-spacepyramid)forfalsepositives(`hardexamples').Allexampleswithscoregreaterthanzeroareconsideredhardexamples.Themethodisthenre-trainedusingthisaugmentedset(initial12180+hardexamples)toproducethefinaldetector.Thesetofhardexamplesissubsampledifnecessary,sothatthedescriptorsofthefinaltrainingsetfitinto1.7GBofRAMforSVMtraining.Startingscaleinscale-spacepyramidaboveisoneandwekeepaddingonemorelevelinthepyramidtillfloor(ImageWidth/Scale)64andfloor(ImageHeight/Scale)128.Scaleratiobetweentwoconsecutivelevelsinthepyramidis1.2.Windowstride(samplingdistancebetweentwoconsecutivewindows)atanyscaleis8pixels.Ifafterfittingallwindowsatascalelevelsomemarginremainsatborders,wedividethemarginby2,takeitsfloorandshiftthewholewindowgrid.Forexample,ifimagesizeatcurrentlevelis(75,130),margin(withstrideof8andwindowsizeof64,128)leftis(3,2).Weshiftallwindowsby(floor(MarginX/2),floor(MarginY/2)).Newimagewidthandheightarecalculatedusingtheformulas:NewWidth=floor(OrigWidth/Scale)andNewHeight=floor(OrigHeight/Scale).Herescale=1impliestheoriginalimagesize.Alsowhiletestingnegativeimages,tocreatenegativewindows,weusethesamesamplingstructure.Youmaydownloadthewholedatasetfromhere(970MB).Toavoidduplicatingimages,'neg'imagefolderin'train_64x128_H96'and'test_64x128_H96'arereferencedusingsymboliclinks.DisclaimerTHISDATASETISPROVIDEDASISANDWITHOUTANYEXPRESSORIMPLIEDWARRANTIES,INCLUDING,WITHOUTLIMITATION,THEIMPLIEDWARRANTIESOFMERCHANTABILITYANDFITNESSFORAPARTICULARPURPOSE.Theimagesprovidedabovemayhavecertaincopyrightissues.Wetakenoguaranteesorresponsibilities,whatsoever,arisingoutofanycopyrightissue.Useatyourownrisk.SupportofEuropeanUnion6thframeworkprojectaceMediaisgreatlyacknowledged.Foranyquestions,commentsorothe