saguar1

YANAI Lab.

電気通信大学 総合情報学科/大学院 総合情報学専攻 メディア情報学コース 柳井研究室
電気通信大学 > 情報工学科 > コンピュータ学講座 > 柳井研究室 > 研究紹介  

Web$BF02h%^%$%K%s%0$N$?$a$N;~6u4VFCD'Cj=P

$BLn8}(B $B82;L(B

2010$BG/(B 2$B7n(B 4$BF|(B




1 $B$O$8$a$K(B

$B1GA|Cf$NF0:n$rG'<1$9$k$3$H$O6K$a$FM-0U5A$G!$MM!9$J%"%W%j%1!<%7%g%s$KMxMQ$9$k$3$H$,$G$-$k$,!$(B $B8=:_%"%N%F!<%7%g%s$5$l$?F0:n$N%G!<%?%;%C%H$OHs>o$K>/$J$$!%$7$+$7!$(B Web$B>e$K$O%?%0$K$h$C$F!$$"$kDxEYJ,N`$5$l$F$$$kF02h$,BgNL$KB8:_$7$F$$$k!%(B $B$h$C$F!$$=$N%?%0$K$h$j1GA|$rBgNL$K<}=8$7!$$=$NJ,N`$r9T$&$3$H$GF0:n$N%G!<%?%;%C%H$r(B $B9=C[$9$k$3$H$,=EMW$K$J$C$F$/$k$H9M$($i$l$k!%(B

$B$=$3$GK\8&5f$G$O!$(BWeb$BF02hJ,N`$KE,$7$?;~6u4VFCD'$rDs0F$7!$$=$NFCD'$rMxMQ$9$k$3$H$G!$ $B$b$&0l$D$NK\8&5f$NL\E*$,B?e$9$k$3$H$,Z$5$l$F$$$k!%K\8&5f$G$O!$F0$-FCD'!$;k3PFCD'!$(B $B;~6u4VFCD'$N(B3$B$D$NFCD'$r(BMultiple Kernel Learning(MKL)$B$K$h$C$FE}9g$7!$F0:nG'<1$r9T$&!%(B $B$J$*K\8&5f$G$OF02h$r%7!<%s%A%'%s%8$N6-3&$GJ,3d$7$?%7%g%C%H$rC10L$H$7$FJ,N`$r9T$&$3$H$H$9$k!%(B

2 $B4XO"8&5f(B

Web$BF02h$K$*$1$kF0:nG'<1$r9T$C$?8&5f$O>/$J$$!%(BCinbis$B$i$O(BWeb$B>e$+$iF0:n$r<+F03X=,$9$k3]$B!%$7$+$7$3$3$GMQ$$$i$l$?!$3X=,%G!<%?$O@EE*$J2hA|$G$"$j!$(B $BCj=P$5$l$kFCD'$b!$@EE*$JFCD'$N$_$rMxMQ$7$F$$$k!%0lJ}$GK\8&5f$G$O!$;k3PE*$JFCD'$N$_$G$O$J$/!$(B $BF0$-FCD'$bG'<1$KMxMQ$9$k!%(B

Liu$B$i$O!$4{B8$N;~6u4VFCD'$H!$;k3PFCD'$rE}9g$9$k$3$H$G(BWeb$BF02h$K$*$1$kF0:nG'<1$r9T$C$F$$$k(B[1]$B!%(B $B$^$?$3$N8&5f$G$O!$(BPage Rank$B$K4p$E$/FCD'A*Br$r9T$&$3$H$G@:EY$r8~>e$5$;$F$$$k!%(B $B$3$N

3 $B8&5fJ}?K(B

$BK\8&5f$G$O!$(B(1)$B;~6u4VFCD'$NCj=P!$(B(2)$BFCD'E}9g$K$h$kF0:nG'<1!$(B(3)$BBgNL%G!<%?$KBP$9$k(BWeb$BF02hJ,N`!$(B $B$N(B3$B$D$N9W8%$,$"$k!%0J2<$G$O$=$l$>$l$N35MW$K$D$$$F=R$Y$F$$$/!%(B

3.1 $B;~6u4VFCD'Cj=P(B

$BK\8&5f$GDs0F$9$k;~6u4VFCD'$O!$(B(1)$B%+%a%i%b!<%7%g%s8!=P!$(B (2)$B;k3PFCD'7hDj!$(B(3)$B;~6u4VFCD'E@$N7hDj!$(B(5)Delaunay$B;03QJ,3d!$(B(4)$BF0$-FCD'$NCj=P!$(B(5)$BFCD'$rE}9g!%(B $B$N(B5$B$D$N%9%F%C%W$KJ,$1$k$3$H$,=PMh$k!%0J2<$G$O$=$l$>$l$K$D$$$F=R$Y$F$$$/!%(B

3.1.0.1 $B%+%a%i%b!<%7%g%s8!=P(B

$BK\8&5f$G$O%+%a%i%b!<%7%g%s$r8!=P$9$k$?$a!$%0%j%C%I>e$KF0$-$rCj=P$7!$0lDj3d9g0J>e$N(B $BNN0h$,F0$$$F$$$?>l9g!$%+%a%i%b!<%7%g%s$N8!=P$H$9$k!%%+%a%i%b!<%7%g%s$,8!=P$5$l$?>l9g(B $B$=$N%U%l!<%`$+$i$OFCD'$OCj=P$5$l$J$$!%(B

3.1.0.2 $B;k3PFCD'$NCj=P(B

$BK\8&5f$G$O!$9bB.$KCj=P$,2DG=$J(BSURF[2]$B$K4p$E$-!$FCD'$rCj=P$9$k!%$3$3$G(B $BCj=P$5$l$?FCD'$O;~6u4VFCD'$N8uJd$H$J$k(B($B?^(B1(1))

3.1.0.3 $B;~6u4VFCD'E@$N7hDj(B

$B;k3PFCD'$G5a$a$i$l$?8uJdE@$NFb!$F0$-$N$J$$E@$O;~6u4VFCD'$H$7$F=EMW$G$O$J$$!%(B $B$h$C$F3FFCD'E@$K4X$7$FF0$-$r7W;;$9$k$3$H$G!$F0$-$N$J$+$C$?E@$r:o=|$7!$(B $BF0$-$N$"$C$?E@$r;~6u4VFCD'E@$H$9$k(B($B?^(B1(2))$B!%(B

3.1.0.4 Delaunay$B;03QJ,3d(B

$BpJs$N$_$G$J$/!$NY@\$7$?E@$N>pJs$b9MN8$7$?FCD'$,9=C[$G$-$k(B($B?^(B1(3))$B!%(B

3.1.0.5 $BF0$-FCD'Cj=P(B

$B?^(B2$B$OF0$-FCD'$NCj=P$NMM;R$r<($7$F$$$k!%A*Br$5$l$?(BN$B%U%l!<%`$r!$(B M$BJ,3d$9$k!%$=$7$F3FFCD'E@$KBP$7$F!$F0$-$r7W;;$9$k$3$H$G!$F0$-$rFCD'2=$9$k!%F0$->pJs$H$7$F$O!$(B $B3FE@$NF0$-$H!$J,3d$5$l$?;03Q7A$NLL@Q$NJQF0$rFCD'2=$9$k$3$H$GF@$i$l$k!%(B $BCj=P$5$l$?FCD'$O(BDollar$B$i$N4]$B$K=>$$!$(Bvideo words$B2=$9$k$3$H$GJ,N`$KMxMQ$5$l$k!%(B

3.2 $BFCD'E}9g$K$h$kF0:nJ,N`(B

$BF0:n$K$h$C$F!$J,N`$9$k$?$a$K=EMW$JFCD'$O0[$J$k!%Nc$($P(B"running"$B$H(B"jogging"$B$rJ,N`$9$k$?$a$K$O!$(B $B;k3PFCD'$G$NJ,N`$OFq$7$$$,!$F0$-FCD'$rMxMQ$9$k$3$H$GJ,N`$r9T$&$3$H$,2DG=$H$J$C$F$/$k!%$=$3$G(B $BK\8&5f$G$O(BMKL$B$K4p$E$-3FFCD'$N=E$_$r<+F0$G7W;;$9$k 3.2.0.1 $B;k3PFCD'(B

6$BJ}8~!$(B4$B<~4|$N%,%\!<%k%U%#%k%?$rMQ$$$F:n@.$5$l$?(B24$B$B$KJ,3d$9$k$3$H$GF@$i$l$k(B$24\times4=96$$B 3.2.0.2 $BF0$-FCD'(B $B%0%j%C%I>e$K%*%W%F%#%+%k%U%m!<$r7W;;$9$k$3$H$GF0$-FCD'$r5a$a$k!%3F%*%W%F%#%+%k%U%m!<$O(B 8$BJ}8~!$(B7$BCJ3,$N%R%9%H%0%i%`$KEjI<$5$l$F$$$/!%(B

3.2.0.3 Bag-of-Frames

$BB?$/$N8&5f$G!$FCD'$O%-!<%U%l!<%`$N$_$+$iCj=P$r9T$C$F$$$k$,!$A*$P$l$k%-!<%U%l!<%`$OG'<1$K(B $BBg$-$J1F6A$rM?$($k!%$7$+$7:G$b=EMW$J%-!<%U%l!<%`$r8+$D$1$k$3$H$OHs>o$K:$Fq$JLdBj$G$"$k!%(B $B$=$3$GK\8&5f$G$O!$(BBag-of-Frames$B$H$$$&?7$?$J9M$($rF3F~$9$k$3$H$G$3$NLdBj$r2r7h$9$k!%(B

Bag-of-Frames$B$H$O!$2hA|G'<1$K$*$1$k(BBag-of-Features$B$rF02h$K3HD%$7$?$b$N$G$"$j!$(B $B%U%l!<%`0lKg$r0l$D$N6I=jFCD'$H8+$J$9$3$H$G!$%U%l!<%`$NFCD'$N=P8=IQEY$GF02hA4BN$rI=8=$9$k(B $B

3.2.0.4 MKL

$BFCD'E}9g$K$O(BMultiple Kernel Learning(MKL)$B$rMxMQ$9$k!%$3$l$OJ#?t$N(BSVM$B%+!<%M%k$r@~7A7k9g$9$k$3$H$K$h$j0l$D$N:GE,$J(B $B%+!<%M%k$r9=C[$9$k
$B?^(B 1: $B;k3PFCD'Cj=P$NMM;R(B
\includegraphics[clip,width=1.0\hsize]{img/visual_change2.eps}
$B?^(B 2: $BA4BN$NF0$-FCD'Cj=P35MW(B($B:8(B)$B!$6I=jDI@W$N35MW(B($B1&(B)
\includegraphics[width=0.7\hsize]{img/motion_method.eps}

$B?^(B 3: KTH$B%G!<%?%;%C%H$K$h$kJ,N`7k2L(B
\includegraphics[width=\hsize]{img/KTH_result2.eps}
$B?^(B 4: KTH$B$K$*$1$k:G?7
\includegraphics[width=\hsize]{img/comp_KTH2.eps}
$B?^(B 5: Our Youtube$B%G!<%?%;%C%H$K$h$kJ,N`7k2L(B
\includegraphics[width=\hsize]{img/youtube_result2.eps}

$B?^(B 6: Wild Youtube$B%G!<%?%;%C%H$NB>
\includegraphics[width=\hsize]{img/wild_comp2.eps}
$B?^(B 7: $B65;U?.9f$"$j$N(BWeb$BF02h%7%g%C%HJ,N`$N7k2L(B
\includegraphics[width=\hsize]{img/web_result2.eps}
$B?^(B 8: k-means$B%/%i%9%?%j%s%0$K$h$k(Bdancing$B$NJ,N`7k2L$NNc(B
\includegraphics[width=\hsize]{img/k-means_dancing2.eps}

   

3.3 $BBgNL(BWeb$BF02hJ,N`(B

$BK\8&5f$G$O!$(BWeb$BF02h$rJ,N`$K$*$$$F!$65;U?.9f$"$j$NJ,N`$K2C$($F!$65;U?.9f$J$7$NJ,N`$b;n$_$?!%(B

3.3.0.1 $B65;U?.9f$"$j(BWeb$BF02h%7%g%C%HJ,N`(B

"walking"$B$H$$$&C18l$G<}=8$5$l$?%7%g%C%H$K$O(B"walking"$B$r4^$`%7%g%C%H$b$"$k$,!$4^$^$J$$%7%g%C%H$b(B $BBgNL$KB8:_$7$F$$$k!%$=$3$GK\ $B$=$l$r9T$&$?$a$K!$<}=8$5$l$?%N%$%:$r4^$`%7%g%C%H$+$i%i%s%@%`$K3X=,%G!<%?$rA*Br$7!$$=$N3X=,%;%C%H$+$i(BSVM$B$r3X=,$9$k!%(B $B$=$N8e!$;D$C$?%F%9%H%;%C%H$KBP$7$F(BSVM$B$GJ,N`$r9T$&!%%i%s%-%s%0$O(BSVM$B$N=PNOCM$K4p$E$$$F9T$o$l$k!%(B

3.3.0.2 $B65;U?.9f$J$7(BWeb$BF02h%7%g%C%HJ,N`(B

"soccer"$B$H$$$&C18l$G<}=8$5$l$?F02h$K$OMM!9$J%7!<%s$,4^$^$l$F$$$k!%$=$3$GK\ $BJ,N`

4 $BI>2A

$BK\8&5f$G$O!$I>2A2A$7$?!%(B

4.1 $B%G!<%?%;%C%H(B

4.1.0.1 $BF0:nG'<1(B

$BF0:nG'<1$NI>2A$r9T$&$?$a!$K\8&5f$G$O(B3$B$D$N%G!<%?%;%C%H$r;HMQ$9$k!%0l$DL\$,!$(BKTH$B%G!<%?%;%C%H$G$"$k!%$3$l$OF0:nG'<1$N8&5f$K$*$$$F(B $B:G$bMxMQ$5$l$F$$$k%G!<%?%;%C%H$G!$(B6$B1]$B!%$3$l$O(B11$B$NF0:n!$7W(B1595$B$N%7%g%C%H$r4^$`!%$3$N%G!<%?%;%C%H$b(B Our Youtube dataset$B$HF1MM$K$h$j@)Ls$N>/$J$$%G!<%?%;%C%H$G$"$k!%(B

4.1.0.2 Web$BF02hJ,N`(B

$B65;U$J$7J,N`$K$O(BOur Youtube datsaset$B$HF1MM$K(B6$B 4.2 $B

4.2.0.1 KTH dataset

$B?^(B3$B$OK\Ds0Fe$9$k$3$H$,J,$+$k!%F0:n$4$H$K8+$?>l9g!$(B"running"$B$H(B"jogging"$B$K$*$$$F(B $B;~6u4VFCD'$h$j!$F0$-FCD'$,J,N`@:EY$,9b$/!$$3$NFs$D$NF0:n$rJ,N`$9$k$?$a$K$OF0$-$,=EMW$G$"$k$3$H$r<($7$F$$$k!%(B

$B$^$??^(B4$B$OK\Ds0F 4.2.0.2 Our Youtube dataset $B?^(B4$B$O(BOur Youtube dataset$B$K$*$1$kJ,N`N($r<($7$F$$$k!%(B $B$?$@$7(BVMR$B$O;~6u4VFCD'$r<($7$F$$$k!%(B $BC10l$NFCD'$G$O!$A4$F$NF0:n$K$*$$$F;~6u4VFCD'$,:G$bNI$$7k2L$G(B77.8%$B$H$J$C$F$$$k!%(B $B$7$+$7FCD'$rE}9g$9$k$3$H$G!$$=$N@:EY$,(B83.3%$B$^$G8~>e$7$F$$$k!%$3$3$G(B "walking"$B$H$$$&F0:n$K$*$$$F!$;k3PFCD'$,F0$-FCD'$r>e2s$C$F$$$k!%$3$l$O;#1F$5$l$k(B"walking" $B$O(B"street"$B$G$"$k2DG=@-$,9b$/!$%7%g%C%H$NGX7J$,N`;w$7$F$$$?$?$a$H9M$($i$l$k!%(B

4.2.0.3 Wild Youtube dataset

$B?^(B6$B$ODs0F1]$B$NHf3S$r<($7$F$$$k!%(B $BK\Ds0Fe$5$;$F$$$k$N$@$,!$$3$NFs$D$NF0:n$O$=$NA*Br$K$h$k(B $B@:EY8~>e$,:G$b82Cx$J$b$N$G$"$C$?$?$a$G$"$k!%K\Ds0Fe$5$;$k$3$H$,=PMh$k$H9M$($i$l$k!%(B

4.2.0.4 $B65;U?.9f$"$j$NBgNL(BWeb$BF02hJ,N`(B

$B?^(B7$B$O65;U?.9f$"$j$N(BWeb$BF02h%i%s%-%s%0$N>e0L(B100$B%7%g%C%H$K$*$1$kE,9gN($r<($7$F$$$k!%$?$@$7(B $B!V%i%s%@%`!W$O%7%g%C%H$r%i%s%@%`$K(B200$BA*Br$7!$$=$NE,9gN($r7W;;$7$?$b$N$G$"$k!%$[$H$s$I$NF0:n$G(B $B;~6u4VFCD'$K$h$k7k2L$,NI$$$,!$(B"eating"$B$K$*$$$F$OF0$-FCD'$NE,9gN($,9b$/$J$C$F$$$k!%(B $B$3$l$O(B"eating"$BFCM-$N>e2<$NF0$-$,6/$/8!=P$5$l$?$?$a$H;W$o$l$k!%$^$?(B"shoot"$B$K$*$$$F$O(B $B>e0L(B100$B$G$OF0$-$H;~6u4VFCD'$,$[$\F1Ey$@$,!$$3$l$O>e0L(B60$B$^$G$r9MN8$9$k$H;~6u4VFCD'$NJ,N`@:EY$,(B $B>!$C$F$$$?!%$7$+$7$=$l0J9_$K$J$k$H!$@52r$N(B"shoot"$B%7!<%s$,8O3i$7$?7k2L(Bmotion$B$HF1Ey$^$GE,9gN($,2<$C$?7k2L$G$"$k$H(B $B9M$($i$l$k!%(B

4.2.0.5 Web$BF02h%7%g%C%H$N65;U?.9f$J$7%/%i%9%?%j%s%0(B

$B?^(B8$B$O(BWeb$BF02h%7%g%C%H$K$*$1$k65;U$J$7%/%i%9%?%j%s%0$N7k2L$N0lNc$r<($7$F$$$k!%$$$/$D$+$N%N%$%:$r(B $B5vMF$9$l$P!$Bg$^$+$J%7!<%sJ,N`$,=PMh$F$$$k$3$H$,J,$+$k!%$7$+$7!$?^(B8$B>eCJ$NCf1{$N%F%m%C%W$N%7%g%C%H$N(B $B$h$&$J%N%$%:$b?tB?$/B8:_$7$F$$$k!%$h$C$FJ,N`$r9T$&A0$K%N%$%:=|5n$r9T$&$3$H$G99$J$k@:EY$N8~>e$,4|BT$G$-$k!%(B

5 $B$*$o$j$K(B

$BK\8&5f$G$O!$(BWeb$BF02hJ,N`$N$?$a$N;~6u4VFCD'Cj=Pl9g!$>e $B:#8e$N2]Bj$H$7$F!$;~6u4VFCD'Cj=P$G$O!$8=>u$G$O%+%a%i%b!<%7%g%s$OGK4~$9$k$3$H$GBP1~$7$F$$$k$,!$(B $B%b!<%7%g%sJd@5$r9T$&$h$&$K$9$k$3$H$,$"$2$i$l$k!%$^$?Cj=P$5$l$?FCD'$+$i=EMW$JFCD'$rA*Br$9$k$3$H$O(B $B6K$a$F=EMW$J$3$H$G$"$k!%BgNL(BWeb$BF02hJ,N`$K4X$7$F$O!$<}=8$9$kF02h$N?t$rA}$d$9!$J,N`A0$K%N%$%:$r=|5n$9$k$3$H$,(B $B$"$2$i$l$k!%(B

$BJ88%L\O?(B

1
J. Liu, J. Luo, and M. Shah.
Recognizing realistic action from videos.
In Proc.of IEEE Computer Vision and Pattern Recognition, pp. 1-8, 2009.

2
B. Herbert, E. Andreas, T. Tinne, and G. Luc.
SURF: Speeded up robust features.
In CVIU, pp. 346-359, 2008.

3
R. I. Cinbins, R. Cinbins, and S. Sclaroff.
Learning action from the web.
In Proc.of IEEE International Conference on Computer Vision, pp. 995-1002, 2009.

4
P. Dollar, G. Cottrell, and S. Belongie.
Behavior recognition via sparse spatio-temporal features.
In Proc. of Surveillance and Performance Evaluation of Tracking and Surveillance, pp. 65-72, 2005.