Click to edit Master text styles
Second level
Third level
Fourth level
Fifth level
‹date/time›
‹foot=
er›
‹#=
250;
------=_NextPart_01C4383C.6E28D720
Content-Location: file:///C:/A2CBB24E/Recognition_files/master04.xml
Content-Transfer-Encoding: quoted-printable
Content-Type: text/xml; charset="utf-8"
Recognizing Objects: Feature Matching
•<=
/span>Problem: Match when viewing conditions change a lot.
–<=
/span>Lighting changes: brightness constancy false.
=
–<=
/span>Viewpoint changes: appearance changes, many viewpoints.
•<=
/span>One Solution: Match using edge-based features.
=
;
–<=
/span>Edges less variable to lighting, viewpoint.
–<=
/span>More compact representation can lead to efficiency=
.
•<=
/span>Match image or object to image
–<=
/span>If object, matching may be asymmetric
<=
/span>
–<=
/span>Object may be 3D.
•<=
/span>Line finding was an example: line=3Dobject; points=
=3Dimage.
![](3D"slide0180_image001.jpg")
![](3D"slide0180_image002.jpg")
Lighting
affects appearance
Problem Definition
•<=
/span>An Image is a set of 2D geometric features, along with positions.
•<=
/span>An Object is a set of 2D/3D geometric =
features, along with positions.
=
•<=
/span>A pose positions the object relative to the image.
–<=
/span>2D Translation; 2D translation + rotation; 2D translation, rotation and scale; planar or 3D obje=
ct positioned in 3D with perspective or scaled orth.&=
#13;
•<=
/span>The best pose places the object features nearest the image features
Strategy
•B=
uild
feature descriptions
•S=
earch
possible poses.
=
–C=
an
search space of poses
<=
/span>
–Or
search feature matches, which produce pose
•T=
ransform
model by pose.
•C=
ompare
transformed model and image.
•P=
ick
best pose.
Presentation Strategy
•Already discussed finding features.=
3;
•First discuss picking =
best
pose since this defines the problem.
•Second discuss search
methods appropriate for 2D.
•Third discuss transfor=
ming
model in 2D and 3D.
•Fourth discuss search =
for
3D objects.
•<=
/span>
•<=
/span>
Example
![](3D"slide0154_image004.jpg")
![](3D"slide0154_image005.gif")
Evaluate Pose
=
226;We
look at this first, since it defines the problem.
&#=
8226;Again,
no perfect measure;
=
211;Trade-offs
between veracity of measure and computational considerations.
Chamfer Matching
![](3D"slide0155_image007.gif")
![](3D"slide0155_image008.gif")
![](3D"slide0155_image009.gif")
![](3D"slide0155_image010.gif")
![](3D"slide0155_image011.gif")
<=
span
style=3D'font-family:Arial;color:#FF9900;mso-color-index:4;mso-special-for=
mat:
lastCR;display:none'>
For every edge point in the transformed object, compute the distance to the nearest image edge point.
Sum distances.
![](3D"slide0155_image013.gif")
![](3D"slide0157_image014.gif")
Main Feature:
• Every model point matc=
hes an
image point.
• An image point can mat=
ch 0,
1, or more model
points.
Variations
•<=
/span>Sum a different distance
–<=
/span>f(d) =3D d2 =
–<=
/span>or Manhatt=
an
distance.
–<=
/span>f(d) =3D 1 if d <
threshold, 0 otherwise.
–This is called bounded
error.
•<=
/span>Use maximum distance instead of sum.
–<=
/span>This is called: directed
Hausdorff distance.
•<=
/span>Use other features
–<=
/span> Corners.
–<=
/span>Lines. Then
position and angles of lines must be similar.
•<=
/span>Model line may be subset of image line.
•
•
•
Other comparisons
•Enforce each image f=
eature
can match only one model feature.
•Enforce
continuity, ordering along curves.
<=
span
class=3DBB style=3D'position:absolute;left:-4.26%'>•These are=
more
complex to optimize.
Presentation Strategy
•Already discussed finding features.=
3;
•First discuss picking =
best
pose since this defines the problem.
•Second discuss search methods appropriate for 2D.
•Third discuss transfor=
ming
model in 2D and 3D.
•Fourth discuss search =
for
3D objects.
•<=
/span>
•<=
/span>
Pose Search
<=
span
class=3DBB style=3D'position:absolute;left:-4.07%'>•Simplest =
approach
is to try every pose.
•Two probl=
ems:
many poses, costly to evaluate each.
•We can re=
duce the
second problem with:![](3D"slide0158_image015.gif")
Pose:
Chamfer Matching with the Distance
Transform
=
0
0
0
0
0
0
0
1
1
1
1
1
1
1
1
1
1
1
1
1
1
2
2
2
2
2
2
2
2
2
2
2
2
3
3
3
3
3
3
3
3
4
=
Example: Each pixel has (Manhat=
tan) distance to nearest edge pixel.=
D.T. Adds Efficiency
![](3D"slide0152_image016.gif")
0
0
0
0
0
0
0
1
1
1
1
1
1
1
1
1
1
1
1
1
1
2
2
2
2
2
2
2
2
2
2
2
2
3
3
3
3
3
3
3
3
4
Then, try all translations of model edges. Add distances under each edge pixel.=
Computing
Distance Transform
•<=
/span>It’s only done once, per problem, not once <=
/span>per pose.
•<=
/span>Basically a shortest path problem.