Download Deformable Templates for Object Recognition: A Statistical Approach - Prof. Brian Potetz and more Study notes Electrical and Electronics Engineering in PDF only on Docsity!
<.=-,/.$>':($?.@)/5A9.$B.+95-.
• C,199.D$E5991353D$")F.3D$G8.5-,/.$!H-/5=I)3$@/)*$85=.2$
J213K$?.@)/5A9.$B.+95-.2DL$MN"0$'OO:P
• 8.97.327Q59AD$GR.+/.2.3-5I)3$53S$?.-.=I)3$)@$
?.@)/*5A9.$#F5+.2DL$6TUM$:;;VP
#,KK.2-.S$R.5S13K
")*+5/13K$-F.$BQ)$B.=F31W,.
• !3./KX$U131175I)3$599)Q2$@)/$)/.$Y.H1A9.$
.3./KX$@,3=I)32P
• TZ$12$K,5/53-..S$-)$[3S$5$K9)A599X$)+I,$
2)9,I)3D$2)$9)3K$52$=)3S1I)32$5/.$*.-P$!3./KX$
131175I)3$=53$K.-$2-,=$13$F)9.2P
• !3./KX$U131*175I)3$235.2$5/.$3)-$5]/5=-.S$-)$
=,/^.2$-F5-$5/.$@5/$5Q5XP
85=159$B/5=\13K$J213K$#35.
&
Part-based models
• Pictorial structures / Constellation of parts
• Rigid parts arranged in a deformable configuration.
- Each part represents local visual properties.
- Spatial configuration captured by statistical model or
spring-like connections.
• Good matching algorithms
- Using dynamic programming and
distance transforms.
template
V
C,199._2$?.@)/5A9.$!X.$B.+95-.
`
C,199._2$?.@)/5A9.$!X.$B.+95-.
a
?.@)/5A9.$B.+95-.$@)/$U5-=F13K
%
• R.W,1/.2$53$131I59$K,.22$-F5-$12$=9)2.$-)$-F.$/.59$)Ab.=-$
• "53$K.-$2-,=$13$9)=59$131
U5-=F13K$A52.S$)3$235.2$53S$K/5S1.3-$S.2=.3-(
B/153K,95-.S$S.@)/5A9.$-.+95-.2$5SS/.22$-F
Deformation model
• Piecewise affine deformations, taking triangles to triangles.
• Measure deformation of template in terms of a sum of
deformation costs per triangle.
• Use a generic deformation model for each triangle, or learn
model using multiple training examples.
template deformation
?.@)/5A9.$B.+95-.$@)/$U5-=F13K
O
• R.W,1/.2$53$131I59$K,.22$-F5-$12$=9)2.$-)$-F.$/.59$)Ab.=-$
• "53$K.-$2-,=$13$9)=59$131
U5-=F13K$A52.S$)3$235.2$53S$K/5S1.3-$S.2=.3-(
B/153K,95-.S$S.@)/5A9.$-.+95-.2$5SS/.22$-F
Deformation model
• Piecewise affine deformations, taking triangles to triangles.
• Measure deformation of template in terms of a sum of
deformation costs per triangle.
• Use a generic deformation model for each triangle, or learn
model using multiple training examples.
template deformation
?.@)/5A9.$B.+95-.$@)/$U5-=F13K
';
B/153K,95-.S$S.@)/5A9.$-.+95-.2$5SS/.22$-F
!3./KX$@,3=I)3$F52$-Q)$+5/-2(
• T$?5-5$-./*$-F5-$.3=),/5K.2$),-./$=)3-),/2$-)$91.$)3$
1*5K.$.SK.2P
• T$6/1)/$-./$-F5-$13F1A1-2$.H=.221^.$S.@)/5I)3P
!c=1.3-$!3./KX$U131*175I)
''
!3./KX$@,3=I)3$F52$-Q)$+5/-2(
• T$?5-5$-./*$-F5-$.3=),/5K.2$),-./$=)3-),/2$-)$91.$)3$
1*5K.$.SK.2P
• T$6/1)/$-./$-F5-$13F1A1-2$.H=.221^.$S.@)/5I)3P
!c=1.3-$!3./KX$U131*175I)
':
!3./KX$@,3=I)3$F52$-Q)$+5/-2(
• T$?5-5$-./*$-F5-$.3=),/5K.2$),-./$=)3-),/2$-)$91.$)3$
1*5K.$.SK.2P
• T$6/1)/$-./$-F5-$13F1A1-2$.H=.221^.$S.@)/5I)3P
B)-59$.3./KX$12$-F.$2,*$)@$.3./K1.2$5-$.5=F$-/153K9.$
E(v) =
T
(EdgeCost(T 1 , T 2 , T 3 ) + W arpCost(T 1 , T 2 , T 3 ))
Figure 11: Matching the corpus callosum model to different images. 19
(a) corpus callosum
(b) maple leaf
10: Models for the corpus callosum (a) and maple leaves (b) generated from binary
es (see text for description). 18
(a) corpus callosum (b) maple leaf
callosum (a) and maple leaves (b) generated from binary
). 18
'O (a) corpus callosum (b) maple leaf
Figure 10: Models for the corpus callosum (a) and maple leaves (b) generated from binary
pictures (see text for description). 18
(a) corpus callosum (b) maple leaf
Figure 10: Models for the corpus callosum (a) and maple leaves (b) generated from binary
pictures (see text for description). 18 Figure 12: Matching the maple leaf model to different images. when it is barely visible. Our matching algorithm performs well in situations where local search techniques tend to fail. To illustrate this we used a public implementation of a local search method known as active appearance models [34]. Every local search technique depends on initialization, so we show results of matching using different initialization parameters on a fixed image. Figure 15 illustrates typical results obtained with active appearance models. It is clear that good initialization parameters are necessary to obtain a good match. This experiment illustrates the advantage of global methods.
4 Learning Deformable Template Models
Now we consider how to learn a deformable shape model for a class of objects from examples.
Intuitively the learning problem is the following. We are given a number of examples for the shape of an object, each of which is a polygon on a fixed number of vertices. Moreover, the vertices of each example are in correspondence with each other. We want to find a triangulated model that can be easily deformed into each of the examples. Each triangle in the model should have an ideal shape and a parameter that controls how much it is allowed to deform. Figure 2 illustrates the learning procedure. Each triangle in the model is shown in its ideal shape, and the triangles are color coded, indicating how much they are allowed to deform. Below we describe how the shape detection problem can be cast in a statistical 20
:; (a) corpus callosum (b) maple leaf
Figure 10: Models for the corpus callosum (a) and maple leaves (b) generated from binary
pictures (see text for description). 18
(a) corpus callosum (b) maple leaf
Figure 10: Models for the corpus callosum (a) and maple leaves (b) generated from binary
pictures (see text for description). 18
:'
Figure 14: Matching to an image corrupted by increasing amounts noise.
Figure 15: Results of a local search method with different initialization parameters. The first row shows the initialization and the second row shows the resulting match in each case. 22
J213K$g/5S1.3-$?.2=.3-
::
Figure 14: Matching to an image corrupted by increasing amounts noise.
Figure 15: Results of a local search method with different initialization parameters. The first row shows the initialization and the second row shows the resulting match in each case. 22
M31I
!2I5-. <)= U131 j)3)+I59$9)=59$1315$5/.$5$2./1),2$+/)A9.( Object recognition • Simple nearest neighbor approach:
- Find minimum deformation necessary to transform object into one of the stored examples. Using a good (shared) deformation model we get high classification accuracy with few training examples. models object deformations
:d
Discussion
• Deformable templates give a simple and compact representation highly variable objects. • We can use a generic deformation model to represent large families of objects using a few examples from each class. • Efficient matching algorithms.
- Exploit structure in classes of models.
- Can find optimal match of model to image quickly.
- Robust to occlusion, noise, etc.
:&
