Learning
Tempo
ralRegularity
inVideo
Seq
uence
Dataand
Cod
es:
http://w
ww.ee.ucr.edu/~mhasan/regularity.htm
l
Mah
mud
ulHasan
1Jong
hyun
Cho
i2JanNeu
man
n2Am
itK.Roy-Cho
wdh
ury1
LarryS.Davis3
Thisresearchispartly
sup
ported
byNSFgrantIIS-13
1693
4and
ONRMURIGrantNo.N00
014-10
-1-093
4.
Anom
alyDe
tection:Com
parison
with
State-of-the
-artM
etho
ds
VisualizingTempo
ralR
egularity
Motivation
qWatchinglonghoursof
uncontrolledvideosisextremely
tedious
Nodatasetbiascompensated
Regularitiesb
ytheGe
neralM
odel
MoreRe
sultsareinth
epa
per
Applicationsand
Experim
ents
Asampleirregularfram
eSynthe
sized
Regularfram
eRe
gularitys
core
Asampleirregularfram
e Syn
thesize
dRe
gularframeRe
gularitys
core
UCSDPe
d2Dataset
UT-Au
stinSu
bway-ExitD
ataset
Pred
ictin
gNearP
asta
ndFuture
CUHK
Avenu
eDa
taset
UT-Au
stinSu
bway-ExitD
ataset
RegularityScores
UT-Au
stinSu
bway-ExitD
ataset
CUHK
Avenu
eDa
taset
UT-Au
stinSu
bway-EnterDataset
FeatureBa
sedFullyCon
nected
Autoe
ncod
er(FCo
nn)
End-to-End
FullyCon
volutio
nalA
utoe
ncod
er(FCo
nv)
UCRiverside1
ComcastLabs,DC2
UniversityofMaryland,CollegePark
3
qWewanttosegment‘m
eaningful’m
omentsinsuch
videoswithoutsupervision
Challenges
qLearningaclassificationmodelofthesemeaningful
(irregular)m
omentsisnottrivialbecause–
qIlldefined(anythingcanbem
eaningful)
qInfrequent(smalltrainingdata)
qLabelingthemisexpensive.
Approa
chq
Usetwohighcapacitygenerativem
odels-deep
neuralnetw
orkbasedauto-encoders(DNN-AE):
qFullyconnectedDNN-AEonhand-craftedfeature
qFullyconvolutionalDNN-AEonframes
qRe
gularinput->Reconstructioncostissmall.
qIrregular
input->Reconstructioncostislarge.
Exe
mpl
ar o
utpu
t of o
ur m
odel
whe
n th
ere
are
irreg
ular
mot
ions
, the
regu
larit
y sc
ore
drop
s sig
nific
antly
.
OurObjectiv
eq
Learningagenerativem
odelforregularitywith
qLimitedsupervisionrequired
qEaseoflearning
qMultipledatasetsusedtotrain
Overviewofthe
App
roach
Mod
elArchitecture
Input:HOG+HOF
collectedaroundthe
trajectoryofinterest
point
Inpu
tDataLayera
ndDataAu
gmen
tatio
nq
Inputcuboidsizes–5,10,and20(Weuse10)
qLargecuboid->betterdiscrim
inationandincreasedrunningtim
e.
qSlidingwindowsize:10,20,and30withsamplerateofstride
1,2,and3respectively.
qSlidingwindowsarem
oved2framesatatim
e.
TrainingaGen
eralM
odel
RegularityScore:s(t)
Optim
izatio
nq
LRScheme:AdaGrad
qInit.LR:0.001(FConv)and
0.01(FConn)
qMini-batchsize:1024
(FConv)and32(FConn)
qWeightinitialization:Xavier
e(x,y,t)=
kI(x,y,t)�
f
W(I(x,y,t))k 2
e(t)=
⌃(x
,y)e(x,y,t)
s(t)
=1�
e(t)�
min
te(t)
max
te(t)