数据集： Middlebury
Vimeo90K
thisdatasetconsistsof 89，800 videoclipsdownloadedfromvimeo.com， whichcoverslargevariatyofscenesandactions.itisdesignedforthefollowingfourvideoprocessingtasks 3360 temporalframeinterpolation volation
tmdhn
with 13320 videos from 101 action categories，tmdhngivesthelargestdiversityintermsofactionsandwithepresenceoflargevariationsincames object appearance and pose、object scale、viewpoint、cluttered background、illumination conditions、etc、 itisthemostchallengingdatasettodate.asmostoftheavailableactionrecognitiondatasetsarenotrealisticandarestagedbyactors、 tmdhnaimstoencouragefurtherresearchintoactionrecognitionbylearningandexploringnewrealisticactioncategories。
thevideosin 101 actioncategoriesaregroupedinto 25 groups，whereeachgroupcanconsistof4-7videosofanaction.thevideosfromthesamege
theactioncategoriescanbedividedintofivetypes :1 (human-object interaction2) Body-Motion Only 3) Human-Human Interaction 4
Adobe-240fps
In total，we collect 71 videos，each with3-5saveragerunningtime.thetrainingvideosarecaptureat 240 fpswithaniphone 6s，goproHero
YouTube-240fps
maybe:http://www.cs.ucf.edu/~ liujg/YouTube _ action _ dataset.html
SlowFlow
微信公众号undervariouslevelsofmotionbluranddifferentmotionmagnitudesillustratedintheanimationbelow。
简单
adatasetfortheevaluationofopticalflowderivedfromtheopensource3danimatedshortfilm
戴维斯