Download Normalized Cuts and Image Denoising: A Multi-scale Approach to Scene Understanding - Prof. and more Study notes Electrical and Electronics Engineering in PDF only on Docsity!
<.=-,/.$>:?($6/)@.=-$AB.
!!"#$%&'$81359$6/)@.=-
- 813B$2)*.-C13D$-C5-$13-./.2-2$E),
- 695E$5/),3B$F1-C$1-
- #..$1G$E),$=53$*5H.$1-$I.J./
- K/1-.$5I),-$1-
- 6/.2.3-$1-$13$= : KC5-$5*$A$/.599E$9))H13D$G)/( !!"#$%&'$81359$6/)@.=-
- !"#$%&'()+"#,%+-%"#).)&&%/'*
- 695E$5/),3B$F1-C$1-
- #..$1G$E),$=53$*5H.$1-$I.J./
- K/1-.$5I),-$1-
- 6/.2.3-$1-$13$= ? KC5-$5$A$/.599E$9))H13D$G)/( "),9B$I.$53$5++91=5L)3M$53$59D)/1-CM$53$1B.5N <)-2$)G$+)-.3L59$-)+1=$1B.52$-)$G)99)FNNN !!"#$%&'$81359$6/)@.=-
- 813B$2)*.-C13D$-C5-$13-./.2-2$E),
- 12-/%-.'0#$%3"+%"**
- #..$1G$E),$=53$*5H.$1-$I.J./
- K/1-.$5I),-$1-
- 6/.2.3-$1-$13$= & KC5-$5$A$/.599E$9))H13D$G)/( A+9..3-$1-M$5H.$1-$F)/HN !!"#$%&'$81359$6/)@.=-
- 813B$2)*.-C13D$-C5-$13-./.2-2$E),
- 695E$5/),3B$F1-C$1-
- 4))%"5%/'0%6-#%(-7)%"%8)9).*
- K/1-.$5I),-$1-
- 6/.2.3-$1-$13$= O KC5-$5$A$/.599E$9))H13D$G)/( P)F$=53$E),$D.-$-C.$I.2-$+./G)/53=.Q P)F$=53$E),$-5E9)/$53$5++/)5=C$-)$5$+5/L=,95/$5++91=5L)3Q R.-$=/.5LS.N$4,-$B)3T-$D.-$/12HEU$#-5/-$21*+9.$V$F)/H$,+F5/B2N !!"#$%&'$81359$6/)@.=-
- K/1-.W,+($XNOY
- Z/59$6/.2.3-5L)3($XNOY
- 6/)@.=-$=)3-.3-($:;Y
- [3B./2-53B13D$)G$-C.$*5-./
- ,591-E$)G$/.2,9-
- "/.5LS1-E
- ]C)/),DC$91-./5-,/.$2.5/=C
- #L=H13D$-)$-C.$+/)+)259$12$.3=),/5D.BM$ #..$*.$1G$E),$D.-$2-,=H (^) ^ R/5B13D( ?OY$)G$E),/$_359$D/5B.
!!"#$%&'$81359$6/)@.=-
- 6/)@.=-$6/)+)2592$`,.($a)3$';b:Xb;%
- K/1J.3$6/)@.=-2$`,.($a)3$':b;'b;%
- Z/59$6/.2.3-5L)32($':b?$W$':b'; X ]1*.913.(
c)$*19.2-)3.$/.+)/-2$5/.$/.d,1/.BN
4,-$1G$E),$D.-$2-,=HM$=).$-59H$-)$.$
$$$e.519b5++)13-.3-b)f=.$C),/2gN
[3.h+.=-.B$C53DW,+2$5/.$S1.F.B$*)/.$
2E+5-C.L=599E$1G$E),$=).$-)$*.$.5/9EN
!!"#$%&'$81359$6/)@.=-
- `,.$a)3$';b:Xb;%
- c)$3..B$-)$F51-$,3L9$a)3B5EN$ A$F199$+/)S1B.$G..BI5=H$53EL*.N
- A3G)/*59N$"),+9.$)G$+5/5D/5+C2N$i),T/.$3)-$
D/5B.B$)3$1-N$]C12$12$-)$C.9+$E),N
- P5S.$2)*.$=)3=/.-.$1B.52N$j3)F$)3.$)/$-F)$
+5+./2$13$-C.$5/.5N$#)*.$+/1)/$/.2.5/=C$F199$C.9+$
E),$B.=1B.$FC5-$122,.2$5/.$C5/B$V$FC5-$
+/)I9.*2$5/.$B)5I9.N
% 6/)@.=-$6/)+)2592( !!"#$%&'$81359$6/)@.=-
- <.-$E),$H3)F$1G$-C.$-)+1=$12$5++/)+/15-.
- <.-$E),$H3)F$1G$-C.$2=)+.$12$)3$-5/D.-
- 6/)S1B.$E),$F1-C$2)*.$/.G./.3=.2$-)$D.-$E),$
2-5/-.BN
- 6)13-$),-$+)221I9.$+1k5992N l A3$/.-,/3M$A$F199( !!"#$%&'$81359$6/)@.=-
- m/),3B$^$+5D.2M$G)/$/.52)35I9.$_D,/.$B.321-E
- 6/.-.3B$1-$12$5$"06n$2,I*1221)
- $)39E$2C)/-./$V$9.22$5*I1L),
- $CJ+(bb=S+/:;;lN)/Db5,-C)/H1-b=S+/H1-N-D
- $`)3T-$2F.5-$G)/*5o3D$B.-5192M$@,2-$G)99)F$-C.$ I521=$2-E9.
- c)F$12$5$D))B$L*.$-)$9.5/3$95-.h$V$I1I-.h
- i),$5/.$/.2+)321I9.$G)/$5$-C)/),DC$91-./5-,/.$2.5/=C '; ]C.$F/1J.3$/.+)/-( !!"#$%&'$81359$6/)@.=-
- 6/)S1B.$5$I/1.G$)S./S1.F$)G$-C.$_.9BN
- 4,-$B)3T-$@,2-$/.D,/D1-5-.$-C.$91-./5-,/.N$]59H$5I),-$ E),/$)F3$F)/HN - #C)F$E),/$/.2,9- - `12=,22$911-5L)32b+/)I9.2$F1-C$-C.$5++/)5=C
- `)3T-$=)S./$.S./E$5-C.5L=59$B./1S5L)3N$j..+$1-$ C1DC$9.S.9N '' ]C.$+/.2.3-5L)3(
';$13,-.$-59H2$F1-C$:W?$13,-.2$G)/$d,.2L)
P1DC$9.S.9$-)+1=2$G/)*$-C.$+/)@.=-2$F199$I.$)3$
-C.$_359$.h5*N
!!"#$%&'$81359$6/)@.=-
- m3E$953D,5D.$ea5-95IM$"ppM$.-=g$12$_3.N
- A+9..3-$E),/$)F3$=)B.N
- [2.$=)B.$G/)$)3913.$)39E$F1-C$E$+./*1221)
- a5-95I$V$A5D.$-))9I)h$5/.$ZjN ': Z-C./$12=.9953.),2$+)13-2(
$8/..53M$6527-)/M$"5/1=C5.9M$q<.5/313D$<)FW<.S.9$0121)3Mr$:;;;N
#13D9.$A*5D.$#,+./Wn.2)9,L)
'l Figure 13. (a) Low-resolution input image. (b) Cubic spline 400% zoom in Adobe Photoshop. (c) Zooming luminance by public domain fractal image compression routine (Polvere, 1998), set for maximum image fidelity (chrominance components were zoomed by cubic spline, to avoid color artifacts). Both (c) and (d) are blurry, or have serious artifacts. (d) Markov network reconstruction using a training set of 10 images taken at the same picnic, none of this person. This is the best possible fair training set for this image. (e) Markov network reconstrution using a training set of generic photographs, none at this picnic or of this person, and fewer than 50% of people. The two Markov network results show good synthesis of hair and eye details, with few artifacts, but (d) looks slightly better (see brow furrow). Edges and textures seem sharp and plausible. (f) is the true full-resolution image. however, the images are ambiguous; Fig. 20 shows dif- ferent scene explanations for a given patch of image data. Both shading and paint scene explanations render to similar image data. We generated 40 such images and their underlying scene explanations at 256 × 256 spatial resolution. Next, given a training image, we broke it into patches, Fig. 19. Because long range interactions are important for this problem, we used a multi-scale ap- proach, taking patches at two different spatial scales, of size 8 × 8 and 16 × 16 pixels. The image patches were sampled with a spatial offset of 7 and 14 pixels, respectively, ensuring consistent alignment of patches across scale, and a spatial overlap of patches, used in computing the compatibility functions for belief prop- agation with Eqs. (8) and (7). As in the other problems, each image patch in the Markov network connects to a node describing the underlying scene variables. For this multi-scale model, each scene node connects to its neighbors in both space and in scale. 4.1. Selecting Scene Candidates For each image patch, we must select a set of candi- date scene interpretations from the training data. For this problem, we found that the selection of candidates required special care to ensure obtaining a sufficiently diverse set of candidates. The difficulty in selecting candidates is to avoid selecting too many similar ones. We want fidelity to the observed image patch, yet at the same time diversity among the scene explanations. A collection of scene candidates is most useful if at least one of them is within! distance of the correct answer. We seek to maximize the probability, P ˆ, that at least one candidate x j i (the^ j th scene candidate at node^ i ) in :;
A*5D.$`.3)1213D
:'
A*5D.$`.3)1213D
::
A*5D.$`.3)1213D
:?
A*5D.$`.3)1213D
:&
A*5D.$`.3)1213D
6)/L995$.-$59M$qImage Denoising using Scale Mixtures of Gaussians$in$the$Wavelet$Domain,” 2003
tral Matting
Rav-Acha 1 Dani Lischinski^1 Eng 2 CSAIL ersity MIT
-
n, at- n r [7]. (a) Input image (b) Hard segmentation (c) Alpha matte (d) Matting components computed by our method. Figure 1. Spectral segmentation and spectral matting This concept is illustrated in Figure 1. Given the input image in Figure 1a, one can produce an unsupervised dis- joint hard partitioning of the image using, e.g., [18] (Figure 1b). In contrast, we compute a set of overlapping, frac- tional, matting components, visualized in Figure 1d. Com- bining three of these components (framed in red) yields the foreground matte of the girl, shown in Figure 1c. In summary, our main contribution is the introduction of the concept of fundamental matting components and the re- sulting first unsupervised matting algorithm. Of course, just like unsupervised segmentation, unsupervised matting is an ill-posed problem. Thus, we also describe two extensions that use these fundamental matting components to construct a particular matte: (i) present the user with several matting alternatives to choose from; or (ii) let the user specify her intent by just a few mouse clicks.
2. Matting components
Matting algorithms typically assume that each pixel Ii in an input image is a linear combination of a foreground color Fi and a background color Bi : Ii =! iFi + ( 1 −! i ) Bi. (1) This is known as the compositing equation. In this work we generalize the compositing equation by assuming that each pixel is a convex combination of K image layers F^1 ,... , FK^ : Ii = K
k = 1 ! ik F ik. (2) 1 :O
#+.=-/59$a5o3D
Levin, Rav-Acha, Lischinski, "Spectral Matting," CVPR 2006. :^
#+.=-/59$a5o3D
CJ+(bbD/519N=2NF52C13D-)3N.B,b+/)@.=-2bd,./Eb 6513-.B e6))/9Eg$2=533.B ]5/D.- Brown University CS143 Intro to Computer Vision ©Michael J. Black Adaboost car detector UIUC Image Database for Car Detection. This database consists of 550 car training images, 500 non-car training images, approximately 170 test images in which cars are present at the same scale as in training, and approximately 170 test images in which cars are present at more than one scale. :X
mB5I))2-$"5/$B.-.=-)/
212 Kutulakos and Seitz Figure 9. Six out of one hundred photographs of a hand sequence. Figure 10 viewpoints. The reconstructed model was computed using an RGB component error threshold of 6%. The model has 112 thousand voxels and. Reconstruction of a hand. An input image is shown in (a) along with views of the reconstruction from the same (b) and other (c)–(f) took 53 minutes to compute. The blue line in (b) indicates the z -axis of the world coordinate system. :% 212 Kutulakos and Seitz Figure 9. Six out of one hundred photographs of a hand sequence. Figure 10 viewpoints. The reconstructed model was computed using an RGB component error threshold of 6%. The model has 112 thousand voxels and. Reconstruction of a hand. An input image is shown in (a) along with views of the reconstruction from the same (b) and other (c)–(f) took 53 minutes to compute. The blue line in (b) indicates the z -axis of the world coordinate system. Kutulakos & Seitz, “A Theory of Shape by Space Carving,” 2000.
#+5=.$"5/S13D
6C)-)$]),/12*$V$6C)-)#E3-C
CJ+(bb91S.95I2N=)*b+C)-)2E3-Cb CJ+(bb+C)-)-),/N=2NF52C13D-)3N.B,b
6C)-)$]),/12*$V$6C)-)#E3-C
CJ+(bb91S.95I2N=)*b+C)-)2E3-Cb CJ+(bb+C)-)-),/N=2NF52C13D-)3N.B,b
