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  

$B0lHL2hA|G'<1$N

1 $BGX7J(B

$B0lHLJ*BNG'<1(B[1]$B$N8&5f$H$O!$FCDj$N@)Ls$N$J$$]$N2hA|$G$bMM!9$J(B $B>uBV$N$b$N$,B8:_$9$k!%$7$+$7!$8=:_$N0lHLJ*BNG'<1$N8&5f$G$O2hA|$N

2 $B8&5fL\E*(B

$BK\8&5f$NL\E*$O!$0lHLJ*BNG'<1$N$?$a$N(B1000$Be$N5pBg2hA|CN<1%G!<(B $B%?%Y!<%9$r9=C[$9$k$3$H$G$"$k!%$=$N$?$a$K!$8=:_Ev8&5f<<$G$O(B1000$Be$N2hA|%G!<%?$,$"$k$,!$$=$N@:EY$O(B$ 40\%$$BDxEY$G$"$k!%$^$?!$$=$N3F%3%s%;%W%H(B $B$K$*$$$F?M $BK\8&5f$G$O!$(B100$B2A$7!$MxMQ2ACM$N9b$$%G!<%?%Y!<%9$N9=C[(B $B$rL\;X$9!%(B

3 $B2hA|G'<1$NJ}K!(B

$B3F%3%s%;%W%HFb$N?M$N%3%s%;%W%H$N@52r(B $B2hA|$+$i%i%s%@%`$K $BFCD'NL$K$O!$6I=jFCD'$N(BSIFT(Scale Inbariant Feature Transform)$BFCD'(B [2]$B$r;HMQ$9$k!%FCD'NL$H$7(B $B$F6I=jFCD'$rMQ$$$k>l9g!$BgNL$NFCD'E@$r=hM}$9$k$?$a!$(Bbag-of-keypoints$B3]$B$rMQ$$$F2hA|$r0l$D$NFCD'%Y%/%H%k$H$7$FI=8=$9$k!%J,N`4o$K$O!$(B SVM(Support Vector Machine)[4] $B$H(BpLSA(Probabilistic Latent Semantic Analysis)[5]$B$rMQ$$$F!$@52r3X=,2hA|$H$NN`;wEY(B $B$r7W;;$9$k!%(B

3.1 Bag-of-keypoints$B

$B6I=jFCD'%Q%?!<%s$N=P8=IQEY(B($B%R%9%H%0%i%`(B)$B$K$h$C$F!$2hA|$rI=8=$9$kJ}K!$G$"(B $B$k!%3F2hA|$+$iB??t$NFCD'E@$rCj=P$7!$3FE@$N6I=j2hA|%Q%?!<%s$r(BSIFT$BK!$G(B128 $B

3.1.0.1 SIFT$BFCD'(B

SIFT$BFCD'$H$O!$FCD'E@<~$j$N6I=j2hA|%Q%?!<%s$r(B128$B!$%9%1!<%kJQ2=!$>HL@JQ2=$KIT(B $BJQ$JFCD'NL$G$"$k!%K\8&5f$G$O3J;RE@FCD'Cj=P$H%i%s%@%`E@FCD'Cj=P$N(B2$B$D$NFC(B $BD'E@Cj=PK!$rMQ$$$F!$(BSIFT$BFCD'$r

3.2 $BJ,N`J}K!(B

3.2.0.1 SVM

SVM$B$O4pK\E*$K(B2$B$D$N%/%i%9$r<1JL$9$kJ,N`4o$r9=@.$9$k$?$a$N3X=,K!$G$"$j!$K\(B $B8&5f$G$O(B $ {\rm SVM}^{light}$[6]$B$rMQ$$$F!$3X=,2hA|$+$iCj=P$7$?(B $BFCD'NL$+$i3X=,%b%G%k$r@8@.$7!$$=$N3X=,%b%G%k$r85$K

3.2.0.2 pLSA

pLSA$B$OE}7W%F%-%9%HJ88%=hM}$+$iH/@8$9$k%b%G%k$G!$3F2hA|$r@x:_%H%T%C%/$N:.(B $B@.$H$7$FI=8=$9$k$l$N2hA|$K$*$1$k(B $B3F%3%s%;%W%H$X$N5"B03NN((B($BN`;wEY(B)$B$r;;=P$9$k!%J8=q(B $ d_{i}(i=1,2,\ldots,I)$$B$K(B $B$*$1$kC18l(B $ w_{j}(j=1,2,\ldots,J)$$B$NH/@8(B $B3NN($r!$@x:_%H%T%C%/(B $ z_{k}(k=1,2,\ldots,K)$$B$rMQ$$$k$H0J2<$N<0$GI=$5$l$k!%(B


$\displaystyle P(w_{j}\vert d_{i})=\sum_{k=1}^{K}P(w_{j}\vert z_{k})P(z_{k}\vert d_{i})$ (1)


$BK\8&5f$G$O!$3X=,%G!<%?$+$i(BpLSA$B$rMQ$$$FF@$i$l$?(B $ P_{train}(d\vert z_{k})$$B$rMQ$$$F!$(B $B3F%H%T%C%/(B$ z_{k}$$B$K5"B0$7$?$b$N$,@52r2hA|$G$"$k3NN((B $ P(OK\vert z_{k})$$B$r7W;;$9(B $B$k!%$^$?!$(B hold-in heuristics[5]$B$N $ P_{test}(z_{k}\vert d_{i})$$B$r5a$a$k!%$3$l$i$NCM$rMQ$$$F!$2hA|(B$ d_{i}$$B$N@52r(B $B3NN((B $ P_{test}(OK\vert d_{i})$$B$r7W;;$9$k$H0J2<$N<0$G$"$i$o$5$l$k!%(B


$\displaystyle P_{test}(OK\vert d_{i})=\sum_{k=1}^{K}P(OK\vert z_{k})P_{test}(z_{k}\vert d_{i})$ (2)


$BK\8&5f$G$O!$@x:_%H%T%C%/$N?t$r(B $ 10,30,50$$B$N(B3$BDL$j$NCM$rMQ$$$F

4 $B

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

$BK\8&5f$G$O!$2hA|%G!<%?%;%C%H$H$7$FEv8&5f<<$G(BWWW$B$+$i<+F0<}=8$7$?2hA|$r;H(B $BMQ$9$k!%$3$3$G$O!$r7o$H$7$F%3%s%;%W%HFb$N<}=8(B $B2hA|Kg?t$,(B1000$BKg0J>e$"$j!$I>2A:QKg?t$,(B45$BKg0J>e$"$k(B100$B]$H$7$F$$$k!%(B $B$3$N(B100$B$B$G$"$k!%(B $BBP>]%3%s%;%W%H$rI=(B1$B$K!$$=$NCf$N(B20$B2A:Q$_@52r2hA|$NNc$r?^(B1$B$K<($9!%(B



$BI=(B 1: 100$B]%G!<%?(B
No. $BC18l(B No. $BC18l(B No. $BC18l(B No. $BC18l(B
1 $B8P(B 26 $BJh(B 51 $BBgJ)(B 76 $BF~3X<0(B
2 $B%@%`(B 27 $BF0J*1`(B 52 $B$&$J$.(B 77 $B<7M<(B
3 $BC+(B 28 $B?eB24[(B 53 $B%+%K(B 78 $B3$?eMa(B
4 $B;3(B 29 $B5e>l(B 54 $B$_$+$s(B 79 $B2V2P(B
5 $B5V(B 30 $B6%GO>l(B 55 $B3A(B 80 $B?@MA(B
6 $B:d(B 31 $BD;5o(B 56 $B$5$L$-$&$I$s(B 81 $B2;3Z2q(B
7 $B2P8}(B 32 $B20Bf(B 57 $B%i!<%a%s(B 82 $BMY$j(B
8 $B?9(B 33 $B8$(B 58 $B%-%c%Y%D(B 83 $B7uF;(B
9 $BEg(B 34 $B%-%D%M(B 59 $B$[$&$l$sAp(B 84 $BAjKP(B
10 $B%S!<%A(B 35 $B>](B 60 $B%H%^%H(B 85 $B%i%0%S!<(B
11 $B5y9A(B 36 $B%Q%s%@(B 61 $BDGB{(B 86 $B%^%i%=%s(B
12 $BN.I9(B 37 $BD;(B 62 $B$7$c$V$7$c$V(B 87 $B%G%9%/(B
13 $B>aF}F6(B 38 $B%M%:%_(B 63 $B%9%F!<%-(B 88 $BBS(B
14 $BKR>l(B 39 $B%/%8%i(B 64 $B>FCq(B 89 $B5m(B
15 $BAR8K(B 40 $B5{(B 65 $B9HCc(B 90 $B5o
16 $BO*E7IwO$(B 41 $B:y(B 66 $B%8%e!<%9(B 91 $B%Z%s(B
17 $B%[!<%k(B 42 $B9HMU(B 67 $B%3!<%i(B 92 $B4$(B
18 $B66(B 43 $B6d0I(B 68 $B%U%'%j!<(B 93 $BMv(B
19 $B$D$j66(B 44 $B>>(B 69 $B%8%'%C%H5!(B 94 $BGO(B
20 $BE4F;(B 45 $B%^%s%0%m!<%V(B 70 $B%T%"%N(B 95 $B7$(B
21 $BNSF;(B 46 $B;gM[2V(B 71 $BB@8](B 96 $B0>(B
22 $B;{(B 47 $B%5%k%S%"(B 72 $B%M%C%/%l%9(B 97 $BAp2V(B
23 $B?@ 48 $B%_%s%H(B 73 $B@c(B 98 $BK_(B
24 $B>k(B 49 $B%5%s%4(B 74 $B1@(B 99 $B@Z
25 $BE7 50 $BV"Iw(B 75 $BM<7J(B 100 $BED?"$((B

$B?^(B 1: 20$B
\includegraphics[width=0.96\hsize]{eps/good-sample3.eps}

4.2 $B

$BBP>]%3%s%;%W%H$NA42hA|$+$i(BSIFT$BFCD'$r3J;RE@!$%i%s%@%`E@$+$i$=$l$>$l $BK\8&5f$G$O!$(B $B2hA|$NFCD'E@Cj=PJ}K!$r3J;RE@!$%i%s%@%`E@$N(B2$BDL$j!$J,N`4o(B $B$r(BSVM$B$H(BpLSA$B$N(B2$B

4.3 $BI>2AJ}K!(B

$BK\8&5f$G$O!$I>2AJ}K!$KL@3N$J5,B'$,$J$$$?$a?M2A$r9T$&I,MW$,$"$k!%$^(B $B$?!$2A$r9T$&$3$H$O:$Fq(B $B$G$"$k!%$=$N$?$a!$3F%3%s%;%W%H$K$*$1$k2A$r9T$$!$$=$N7k2L$r%3%s%;%W%HA4BN$NI>2A$H$7$FMxMQ$9$k%i(B $B%s%@%`%5%s%W%j%s%0$N $BJ,N`7k2L$NI>2A$KMQ$$$k4p=`$H$7$F!$E,9gN($GJ,N`%7%9%F%`$N@-G=$rI>2A$9$k!%(B $B3F%3%s%;%W%H$K4^$^$l$k@52r2hA|$NKg?t$O!$%i%s%@%`%5%s%W%j%s%0$K$h$C$FF@$i(B $B$l$?3:Ev%3%s%;%W%HA4BN$NE,9gN($rMQ$$$F5a$a$?!%(B
$B$3$3$G!$E,9gN($O

$\displaystyle {\rm Precision}($$BE,9gN((B$\displaystyle )= \tfrac{{\rm True\ Positive}}{{\rm True\ Positive}+{\rm False\ Positive}}$ (3)


$BI>2A$O!$N`;wEY$N>e0L(B1000$B0L$G$NE,9gN($H!$3F%3%s%;%W%H$K4^$^$l$k@52r2hA|Kg(B $B?t$G$NE,9gN((B $B$r5a$a$?!%$^$?!$%3%s%;%W%HA4BN$G$N(B11$BE@J?6QE,9gN($b5a$aI>2A$r9T$C$?!%(B11$BE@(B $BJ?6QE,9gN((B$ \tilde{P}$$B$O:F8=N(%l%Y%k(B $ x(x=0.0,0.1,\ldots,1.0)$$B$K$*$1$kE,9g(B $BN((B$ P(x)$$B$r(B $BMQ$$$F0J2<$N$h$&$K7W;;$9$k$3$H$,$G$-$k!%(B


$\displaystyle \tilde{P}=\frac{1}{11}\sum _{i=0}^{10}P(\frac{i}{10})$ (4)


4.4 $B

$B%i%s%@%`E@FCD'Cj=P$r9T$$!$(BSVM$B$rMQ$$$?>l9g$N7k2L$r<($9!%$3$3$G!$(B $B?^(B2$B$O!V8$!W$N%3%s%;%W%H$K$*$1$k%i%s%/IU$1$r9T$C$?>e0L(B100$B0L(B $B$H2<0L(B100$B0L$K4^$^$l(B $B$k2hA|$NNc$G$"$k!%@D?'$O(BGOOD$B2hA|!$NP?'$O(BOK$B2hA|!$@V?'$O(BNG$B2hA|!$3%?'$OL$I>(B $B2A2hA|$G$"$k!%(B
$BI=(B2$B$K!$N`;wEY$N>e0L(B1000$B0L$G$NE,9gN($NJ?6Q!$@52r2hA|Kg?t(B $B$^$G$NE,(B $B9gN($NJ?6Q!$(B11$BE@J?6QE,9gN($NJ?6Q$r<($9!%(B
$B0J>e$N7k2L$+$i!$%i%s%@%`E@FCD'Cj=P$rMQ$$$F!$(BSVM$B$GJ,N`$r9T$C$?>l9g$N7k2L$,(B 1000$B0L$^$G$NE,9gN($NJ?6Q!$@52r2hA|Kg?t$^$G$NE,9gN($NJ?6Q!$(B11$BE@J?6QE,9gN((B $B$NJ?6Q$K$*$1$kCM$G!$:G$bNI$$7k2L$H$J$C$?!%$3$N7k2L$rMQ$$$F>e0L(B1000$B0L$^$G(B $B$NE,9gN($+$i(B100$B$B$+$i!$(B$ 3.5$$B%]%$%s%H$NA}2C$,$_$i$l$?!%(B
$B$^$?!$A4%3%s%;%W%H$KBP$7$F(BSVM$B$NJ}$,NI$$7k2L$H$J$C$?$o$1$G$O$J$/!$(B $B?^(B3(a)$B$K<($9$h$&$K!$!VJh!W$G$O(BpLSA$B$G$NJ,N`7k2L$NJ}$,NI(B $B$$7k2L$H$J$k>l9g$b$"$C$?!%(B

$B?^(B 2: $B!V8$!W$G$NJ,N`7k2L$NNc(B($B%i%s%@%`E@(BSVM)
\includegraphics[width=0.6\hsize]{eps/pos100-2.eps} \includegraphics[width=0.4\hsize]{eps/neg100-2.eps}
(a)$B%i%s%-%s%0>e0L(B (b)$B%i%s%-%s%02<0L(B



$BI=(B 2: $B3FJ,N`4o$K$*$1$kE,9gN($NJ?6QCM(B
$BFCD'E@(B 1000$B0L$^$G$N(B $B@52r2hA|Kg?t$^$G$N(B 11$BE@J?6QE,9gN((B
$BJ,N` $BCj=P $BE,9gN($NJ?6Q(B(%) $BE,9gN($NJ?6Q(B(%) $B$NJ?6Q(B(%)
SVM $B3J;RE@(B 47.046 54.356 57.989
SVM $B%i%s%@%`(B 47.511 55.104 58.748
pLSA(10) $B3J;RE@(B 45.706 51.515 55.260
pLSA(30) $B3J;RE@(B 46.377 52.897 56.814
pLSA(50) $B3J;RE@(B 46.430 52.684 57.066
pLSA(10) $B%i%s%@%`(B 43.636 47.679 51.597
pLSA(30) $B%i%s%@%`(B 45.156 51.395 54.534
pLSA(50) $B%i%s%@%`(B 46.137 51.863 56.058

$B?^(B 3: SVM$B$H(BpLSA$B$NE,9gN($NHf3S(B
\includegraphics[width=0.5\hsize]{eps/haka-result.eps} \includegraphics[width=0.5\hsize]{eps/tomato-result.eps}
(a)$B!VJh!W$N>l9g(B (b)$B!V%H%^%H!W$N>l9g(B

5 $B9M;!(B

pLSA$B$G$NJ,N`$K$*$$$F!$@x:_%H%T%C%/?t$O;vA0$K7h$a$kI,MW$,$"$k!%$=$N$?$a!$(B $BK\8&5f$G$O@x:_%H%T%C%/?t$r(B$ 10,30,50$$B$N(B3$BDL$j$GJ,N`$r9T$C$?!%$=$N7k2L!$(B $BI=(B2$B$+$i$o$+$k$h$&$K!$@x:_%H%T%C%/?t$,Bg$-$$J}$,7k2L$,(B $BNI$$!%$7$+$7!$@x:_%H%T%C%/?t$,Bg$-$/$J$l$P$J$k$[$I!$J,N`$H4X78$N$J$$(B $B%H%T%C%/$^$G$bG'<1BP>]$H$7$F$7$^$$!$@:EY$NDc2<$K$D$J$,$k62$l$,$"$k$?$a!$(B $B$=$l$>$l$N%3%s%;%W%H$K$*$1$k:GE,$J@x:_%H%T%C%/?t$r5a$a$kI,MW$,$"$k$H(B $B9M$($i$l$k!%(B
$B$^$?!$(BSVM$B$H(BpLSA$B$NJ,N`@:EY$rHf3S$7$?7k2L!$(BpLSA$B$h$j$b(BSVM$B$NJ}$,J,N`@:EY$N9b$$(B $B>l9g$,B?$/!$(BSVM$B$,2hA|J,N`$K$*$$$FM-8z$J2$B$+$i8+$F

6 $B:#8e$N2]Bj(B

$B<}=8Kg?t$N>/$J$$%3%s%;%W%H$K4X$7$F!$$5$i$K2hA|<}=8$r9T$$8=%G!<%?(B $B%;%C%H$r3HD%$7$F$$$/I,MW$,$"$k!%$^$?!$3F%3%s%;%W%H$K$*$1$kI>2A:Q(B $B$_2hA|$K$D$$$F$b!$8=%G!<%?%;%C%H$G$O;wDL$C$?$b$N$P$+$j$G@52r2hA|%G!<%?$r9=@.(B $B$7$F$$$k$b$N$,$"$C$?!%$=$N$?$a!$I>2A:Q$_2hA|$N:F9=C[$r9T$$!$B?MM@-$N$"$k@52r(B $B2hA|%G!<%?$K$7$F$$$+$J$1$l$P$$$1$J$$!%(B
$B$^$?!$%^%k%A%/%i%9J,N`$r9T$&$3$H$G!$%3%s%;%W%H$NB?MM@-$K$D$$$F$b:#8eBP1~$7(B $B$F$$$/$3$H$,=EMW$G$"$k$H9M$($i$l$k!%$=$N$?$a!$(B $B:#2s7k2L$NNI$+$C$?(BSVM$B$@$1$rMQ$$$k$N$G$O!$%^%k%A%/%i%9J,N`$KE,$7$?J,N`Z$7$F$$$/I,MW$,$"$k!%(B
$BB>$K$b!$K\8&5f$G9=C[$7$?2hA|%G!<%?%Y!<%9$K$O%N%$%:$H$J$kIT@52r2hA|$,F~$C$F(B $B$$$k$?$a!$%N%$%:2hA|$N=|5nJ}K!$K$D$$$F$b:#8e3NN)$7$F$$$/I,MW$,$"$k!%(B


$BJ88%L\O?(B

1
$BLx0f7<;J(B.
$B0lHLJ*BNG'<1$N8=>u$H:#8e(B.
$B>pJs=hM}3X2qO@J8;o(B: $B%3%s%T%e!<%?%S%8%g%s!&%$%a!<%8%a%G%#%"(B, Vol. 48, No. SIG16 (CVIM19), pp. 1-24, 2007.

2
D. G. Lowe.
Distinctive Image Features from Scale-Invariant Keypoints.
International Journal of Computer Vision, Vol. 60, No. 2, pp. 91-110, 2004.

3
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.

4
N. Cristianini and J. Shawe-Taylor.
$B%5%]!<%H%Y%/%?!<%^%7%sF~Lg(B.
$B6&N)=PHG(B, 2005.

5
T. Hofmann.
Unsupervised Learning by Probabilistic Latent Semantic Analysis.
Machine Learning, Vol. 43, pp. 177-196, 2001.

6
Thorsten Joachims.
$ {\rm svm}^{light}$.
http://www.cs.cornell.edu/People/tj/svm_light/index.html.