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  

 

ホーム

研究内容    【概要】    【研究紹介】

メンバー

発表文献

場所

アルバム (学内)

内部情報 (研究室内)

 

卒研配属情報 (学内)

 

English Page

$BBgNL2hA|$N0UL#E*J,N`(B

$B=)4V(B $BM:B@(B

2008$BG/(B 2$B7n(B 7$BF|(B

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

$B$3$l$^$G$N2hA|J,N`$G$O!$308+E*FCD'$rMQ$$$F4X78@-$NGv$$2hA|72$KBP$7$F$N(B $B%+%F%4%jJ,N`$, $B$=$3$G!$$5$i$KG'<1$dJ,N`$r>\$7$/9T$&$?$a$K!$0UL#E*$K0lCW$7$F$$$k2hA|$r(B $B$5$i$K?<$$0UL#3,AX$GJ,N`$9$kI,MW@-$,$"$k!%6aG/$G$O!$308+E*FCD'$r(B $BMQ$$$?2hA|$N0UL#E*J,N`$rBj:`$K$7$?O@J8$bH/I=$5$l$F$$$k$,!$?t$b>/$J$/!$(B $BI,$:$7$bNI$$7k2L$,F@$i$l$F$$$k$o$1$G$O$J$$!%(B

$B8&5f$N:G=*E*$JL\E*$H$7$F$O!$BgNL2hA|$r0UL#$N$"$k3,AX$GJ,N`$7$F$$$/$3$H$G$"$k$,!$(B $B$=$N$?$a$K$O!$$"$kDxEY$N@:EY$G2hA|$r0UL#E*$KJ,N`$9$kI,MW$,$"$k$H9M$($i$l$k!%(B $B$=$3$GK\8&5f$G$O!$MM!9$Je$rL\;X$7$?!%(B

2 $B2hA|J,N`$NJ}?K(B

$BK\h$jJ*!W$N(B2$B$D$N%+%F%4%j$K$D$$$F$N0UL#E*J,N`$r9T$&!%(B $B$=$l$>$l%+%F%4%jFb$N(B10$B $BJ,N`$r9T$&A0$K!$;w$?2hA|$NJ,N`$G$O2hA|Fb$NBP>]$H$J$kItJ,0J30$O(B $BJ,N`$N $BJ,N`$O!$(BpLSA$B!$(BpLSA$B%Y%/%H%k!$(Bbag-of-keypoints$BI=8=$GI=$7$?FCD'%Y%/%H%k(B($B%Y!<%9%i%$%s(B)$B$r;H$C$F9T$&!%(B

3 $BJ,N`$KMQ$$$k

3.1 $B2hA|FCD'Cj=P(B

3.1.1 $B6I=jFCD'(B

3.1.1.1 SIFT$BFCD'(B

SIFT(Scale Invariant Feature Transformation)$BFCD'$O(B David Lowe$B$K$h$C$FDs0F$5$l!"FCD'E@$N8!=P(B(SIFT$BFCD'(B)$B$H(B128$B!$%9%1!<%kJQ2=$K6/$$!%(B

3.1.1.2 Color SIFT

$B2hA|$r(BHSV$B?'6u4V$KJQ49$7$?8e!$?'Aj(B(H)$B!$:LEY(B(S)$B!$L@EY(B(V)$B$N$=$l$>$l$G(BSIFT$B5-=R;R$r(B $B;;=P$7!$$=$l$i$rBP1~$7$?E@$GE}9g$9$k$3$H$GF@$i$l$?E@$rFCD'E@$H$9$k$b$N$G$"$k!%(B $B$=$N$?$a!$(BColor SIFT$B5-=R;R$O(B$128\times3$$B

3.1.2 Bag-of-keypoints

Bag-of-keypoints$B%b%G%k(B[1]$B$O!$AjBP0LCV$rMQ$$$:$K2hA|$r6I=jFCD'$N=89g$HB*$($?(B $B

3.1.3 pLSA$B%Y%/%H%k(B

pLSA$B$rMQ$$$k$3$H$G!$2hA|(B$d$$B$,3F%H%T%C%/(B$z$$B$K=jB0$9$k3NN((B$P(z\vert d)$$B$,5a$^$j!$(B $B%H%T%C%/?t$N$B$H$9$k$HA4$F$N2hA|$r(B $K$$B2]$B$H8F$V$3$H$H$9$k!%(B pLSA$B$N>\:Y$O(B3.3$B$G=R$Y$k!%(B

3.2 $BGX7J=|5n(B

$BGX7J=|5n$O!$BP>]$NFCD'$G$"$k3NN((B$P(fg\vert w)$$B$r5a$a$F!$$=$NCM$,$"$k0lDjCM$h$j9b$$FCD'$r(B $BBP>]$NFCD'$H$7$F;HMQ$9$k!%BP>]$NFCD'$G$"$k3NN((B$P(fg\vert w)$$B$N5a$aJ}$H$7$F(B2$B 3.2.1 $BGX7J%G!<%?%;%C%H(B $BGX7J%G!<%?%;%C%H$O!VD;!W!$!V>h$jJ*!W$N$=$l$>$l$N%+%F%4%j$GMQ0U$7!$!VD;!W$K$OAp!$8P!$6u!$LZ$N;^$r!$(B $B!V>h$jJ*!W$K$O3$!$7zJ*!$@~O)!$F;!$6u!$@c86!$EZ!$LZ$rGX7J2hA|$Ne5-$N 3.2.2 $B=|5nJ}K!(B1 : $BEjI<(B $BGX7J%G!<%?%;%C%H$HBP>]2hA|72$r%/%i%9%?%j%s%0$7$FF@$i$l$?(Bvisual word$B$+$i!$(B $B3F2hA|$KBP$7$F!$$=$N(Bvisual word$B$,H/8+$5$l$?2hA|$,GX7J%G!<%?%;%C%H$N2hA|$G$"$l$P!$(B $BGX7J$NFCD'$H$7$FEjI<$7!$$=$l0J30$J$i!$BP>]2hA|$NFCD'$H$7$FEjI<$7!$:G=*E*$K(B $BBP>]$H$7$FEjI<$5$l$?3d9g$r3NN((B$P(fg\vert w)$$B$H$9$k!%(B

3.2.3 $B=|5nJ}K!(B2 : pLSA$B$G5a$^$C$?CM$+$i$N;;=P(B

$BGX7J%G!<%?%;%C%H$HBP>]2hA|72$r%/%i%9%?%j%s%0$7$FF@$?(Bvisual word$B$+$i!$(Bbag-of-keypoints$BI=8=$r(B $B9T$$!$(BpLSA$B$KE,1~$7$F5a$^$k%+%F%4%j(B$z$$B$KBP$7$F2hA|(B$d$$B$,4^$^$l$k3NN((B$P(d\vert z)$$B!$(Bvisual word$B$G$"$k(B $w$$B$,H/@8$7$?$H$-$K$I$N%+%F%4%j(B$z$$B$G$"$k$+$N3NN((B$P(z\vert w)$$B$r;H$C$F!$(Bvisual word$B$,BP>]2hA|$N(B $BFCD'$G$"$k3NN((B$P(fg\vert w)$$B$r5a$a$k!%(B

3.3 pLSA

pLSA(probabilistic Latent Semantic Analysis)[3]$B$O$b$H$b$H%F%-%9%H=hM}$KBP$7$F9M$($i$l$?=hM}$G!$J8=q$HC18l!$(B $BC18l$+$i?dB,$5$l$kOCBj$K4X$9$k4X78$r3NN(E*$K;;=P$9$k $d\in D = \{d_1,...,d_N\}$$B$H$7$F2hA|!$C18l(B $w\in W = \{w_1,...,w_M\}$$B$H$7$F(Bvisual word$B!"OCBj(B $z\in Z = \{z_1,...,z_K\}$$B$O%+%F%4%j$H$7$FBP1~$5$;$?!%(B

$B<0(B(1)$B$N4X78$+$i!$(BEM$B%"%k%4%j%:%`$K$h$jCM$N?dDj$,$G$-$k!%(B

$$P(w_i\vert d_j) = \sum^K_{k=1} P(w_i\vert z_k)P(z_k\vert d_j)$$ (1)

4 $B

4.1 $B

$B?^(B 1: $Be$+$i!$(Bbald eagle$B!$(Bosprey$B!$(B red-tailed hawk$B!$(Bowl$B!$(Bmallard$B!$(Bduckling$B!$(Bmacaw$B!$(Bheron$B!$(Bfinch$B!$(Bpuffin$B!%(B $B1&$O!V>h$jJ*!W$G:8>e$+$i!$(Bskidoo$B!$(Bmotorbike$B!$(Bmountain bike$B!$(Bbullet train$B!$(Bsail boat$B!$(Bbuses$B!$(Bsports car$B!$(Bhover craft$B!$(Bwhirly bird$B!$(Bjetliner$B!%(B)

$B$3$3$G$N!$L\E*$O!VD;!W!$!V>h$jJ*!W$=$l$>$l$N%+%F%4%j(B($B?^(B1)$B$G2hA|J,N`$r9T$&$3$H$G$"$k!%(B $B$^$?!$$=$l$>$l$N%+%F%4%j$N(B10$Bl9g(B($B?^(B1$B$G$NF1$8?'$N%0%k!<%W(B)$B$NJ,N`$b$*$3$J$C$?!%(B

$B$^$?!$J,N`$N@:EY$r$"$2$k$?$a$K!$GX7J=|5n$N $B2hA|J,N`$NJ}K!$OI=(B1$B$N(B3$BDL$j$G$"$k!%(B

$BI=(B 1: $B2hA|J,N`J}K!(B

plsa	pLSA$B$K$h$k2hA|(B$d$$B$,%H%T%C%/(B$z$$B$K(B
	$B=jB0$9$k3NN((B$P(z\vert d)$$B$G$NJ,N`(B
plsa vector	pLSA$B%Y%/%H%k(B(pLSA$B$K$h$k2hA|I=8=(B)$B$r(B
	k-means$B$GJ,N`(B
base line	bag-of-keypoints$BI=8=$G(B
	$BI=$7$?FCD'%Y%/%H%k$r(Bk-means$B$GJ,N`(B

4.2 $BI>2AJ}K!(B

$B2hA|J,N`$NI>2A$OJ,N`N($G9T$&$3$H$H$7$?!%3F%/%i%9%?(B$c$$B$G:G$b?t$NB?$$$B$H$9$k$H!$(B $BJ,N`N($O(B $\frac{\sum n_c^{max}}{\sum N_c}$$B$G5a$a$i$l$k!%(B

4.3 $B

4.3.1 $BGX7J=|5n$N7k2L(B

$B:G$bGX7J$N=|5n$,$&$^$/8@$C$F$$$k$H;W$o$l$k7k2L$r?^(B2$B$K<($7$?!%(B $B2hA|J,N`$GGX7J$N=|5n$r9T$&$H$-$K$O!$?^(B2$B$N$H$-$HF1MM$N@_Dj$r;H$&$3$H$H$7$?!%(B

$B?^(B 2: $BBP>]$NFCD'Cj=P$N7k2L(B

4.3.2 $B2hA|J,N`$N7k2L(B

$B?^(B3$B$G!$%/%i%9%??t$r(B60$B$K8GDj$7$?$H$-$N3F 10$Bh$jJ*%+%F%4%j$G(Bk-means$B$K$h$kJ,N`$N$H$-J,N`N($,:GBg$N(B41%$B$H$J$C$?!%(B

$B?^(B 3: $B3Fe$,GX7J=|5n$J$7!$2<$,GX7J=|5n$"$j(B)

4.4 $B9M;!(B

4.4.1 $BGX7J=|5n(B

$B:#2s$N]$NFCD'$N5a$aJ}$N0c$$$G$O!$$=$l$[$I=|5n$N7k2L$K:9$O8=$l$J$+$C$?!%(B

$B0lHV!$GX7J$N=|5n$,0lHV$&$^$/$$$C$F$$$?$N$O(BColor SIFT$B$r;H$C$F!$L@EY$J$I$K(B $B]$NFCD'$,0lCW$7$F$7$^$&$N$rKI$.!$GX7J%G!<%?%;%C%H$NL@EY$rJQ$($k$3$H$G!$(B $B2hA|$NGX7J$N:Y$+$$JQ2=$KBP1~$G$-$F$$$k$+$i$G$"$k$H;W$o$l$k!%(B

4.4.2 $B2hA|J,N`(B

$B3$B$+$i!$:#2s07$C$? $B$^$?!$:#2s$N3$B$+$i$o$+$k!%$3$l$OGX7J$N>pJs$,J,N`$KM-8z$G$"$k$3$H$r<($9!%(B $B$?$$$F$$$N@8J*$J$I$O!$$=$N$*$+$l$F$$$k4D6-$KE,$9$k$h$&$K$G$-$F$$$k$N$G!$$^$:$=$&$$$C$?4D6-$G(B $BJ,N`$7$F$+$i!$$=$N4D6-Fb$NBP>]$GJ,N`$9$k$3$H$,M-8z$K$J$k$3$H$,9M$($i$l$k!%(B

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

$BK\8&5f$G$O!$0UL#E*J,N`$r@.$7?k$2$k$?$a$K!$(BpLSA$B$dGX7J=|5n$J$I$NpJs$N=EMW@-$r3NG'$G$-$?$3$H$OHs>o$K0UL#$N$"$k$3$H$G$"$k$H(B $B;W$o$l$k!%(B

$B$7$+$7!$2hA|$r0UL#E*$KJ,N`$9$k$3$H$,$?$$$X$s:$Fq$G$"$k$H$$$&$3$H$r2~$a$F $BK\8&5f$G!$(BpLSA$B%Y%/%H%k$rMQ$$$?$N$OJ,N`$r9T$&$?$a$@$1$G$J$/!$3,AXE*$KJ,N`$9$k$3$H$r;kLn$KF~$l$?$b$N$G$"$C$?!%(B $B$=$l$O!$(BpLSA$B%Y%/%H%k$rMQ$$$k$3$H$G2hA|$N%(%s%H%m%T!<$r;;=P$9$k$3$H$,2DG=$G$"$j!$%(%s%H%m%T!

$BJ88%L\O?(B

1

G. Csurka, C. Bray, C. Dance, and L. Fan.
Visual categorization with bags of keypoints.
In Proc. of ECCV Workshop on Statistical Learning in Computer Vision, pp. 1-22, 2004.

2

$B9>ED5#@2(B, $B5H@n@5=S(B, $B;3<<2m;J(B.
Folksonomy $B$N%?%0$rMQ$$$?<+F0J,N`BN7O9=C[$X8~$1$F(B ($B%/%i%9%?%j%s%0!&%+%F%4%i%$%:(B, $B2F$N%G!<%?%Y!<%9%o!<%/%7%g%C%W(B 2007 ($B%G!<%?9)3X(B, $B0lHL(B)).
$BEE;R>pJsDL?.3X2q5;=Q8&5fJs9p(B. DE, $B%G!<%?9)3X(B, Vol. 107, No. 131, pp. 405-410, 2007.

3

T. Hofmann.
Unsupervised learning by probabilistic latent semantic analysis.
Machine Learning, Vol. 43, pp. 177-196, 2001.