Lecture 06

Computer Vision: Image Alignment
Raquel Urtasun
TTI Chicago
Jan 24, 2013
Raquel Urtasun (TTI-C)
Computer Vision
Jan 24, 2013
1 / 44
Readings
Chapter 2.1, 3.6, 4.3 and 6.1 of Szeliskis book

Chapter 1 of Forsyth & Ponce
Computer Vision
Jan 24, 2013
2 / 44
What did we see in class last week?
Computer Vision
Jan 24, 2013
3 / 44
What is the geometric relationship between these images?
!"
[Source: N. Snavely]
Computer Vision
Jan 24, 2013
4 / 44
What is the geometric relationship between these images?
Very important for creating mosaics!

Computer Vision
Jan 24, 2013
4 / 44
Image Warping
Image filtering: change range of image
g (x) = h(f (x))
!"
%"
$"
#"
#"
Image warping: change domain of image

g (x) = f (h(x))
!"
%"
$"
#"
#"
[Source: R. Szeliski]
Computer Vision
Jan 24, 2013
5 / 44
Parametric (global) warping
!"
!!"!#$%&'!
!"!"!#$(%&('!
Transformation T is a coordinate-changing machine:

p 0 = T (p)
What does it mean that T is global?
Is the same for any point p
Can be described by just a few numbers (parameters)
Computer Vision
Jan 24, 2013
6 / 44
Forward and Inverse Warping

Forward Warping: Send each pixel f (x) to its corresponding location
(x 0 , y 0 ) = T (x, y ) in g (x 0 , y 0 )
Inverse Warping: Each pixel at the destination is sampled from the original
image
Computer Vision
Jan 24, 2013
7 / 44
All 2D Linear Transformations
Linear transformations are combinations of

Scale,
Rotation
Shear
Mirror
0
x
a
=
y0
c
b
d

x
y
Computer Vision
Jan 24, 2013
8 / 44

Properties of linear transformations:
Origin maps to origin
Lines map to lines
Computer Vision
Jan 24, 2013
9 / 44

Lines map to lines
Parallel lines remain parallel
Computer Vision
Jan 24, 2013
9 / 44

Lines map to lines
Ratios are preserved
Computer Vision
Jan 24, 2013
9 / 44

Lines map to lines
Closed under composition
0
x
a
=
c
y0
b
d

e
g
f
h

Computer Vision
i
k
j
l

x
y
Jan 24, 2013
9 / 44

Lines map to lines
0
x
a
=
c
y0
b
d

e
g
f
h

i
k
j
l

x
y
What about the translation?

Computer Vision
Jan 24, 2013
9 / 44

Lines map to lines
0
x
a
=
c
y0
b
d

e
g
f
h

i
k
j
l

x
y
What about the translation?

Computer Vision
Jan 24, 2013
9 / 44
Affine Transformations
Affine transformations are combinations of
Linear transformations, and
Translations
0
a
x
y 0 = d
w
0
b
e
0

c
x
f y
1
w
Properties of affine transformations:

Origin does not necessarily map to origin
Lines map to lines
Computer Vision
Jan 24, 2013
10 / 44
Projective Transformations
Affine transformations and Projective
0
a
x
y 0 = d
w0
g
warps
b
e
h

c
x
f y
i
w
Properties of affine transformations:

Origin does not necessarily map to origin
Lines map to lines
Parallel lines do not necessarily remain parallel
Ratios are not preserved
Computer Vision
Jan 24, 2013
11 / 44
2D Image Tranformations
These transformations are a nested set of groups

Closed under composition and inverse is a member
Computer Vision
Jan 24, 2013
12 / 44
Computing transformations
Given a set of matches between images A and B
How can we compute the transform T from A to B?
Find transform T that best agrees with the matches
Computer Vision
Jan 24, 2013
13 / 44
Least squares formulation

For each point (xi , yi ) we have
xi + xt
yi + yt
=
=
xi0
yi0
We define the residuals as

rxi (xt ) = xi + xt xi0
ryi (yt ) = yi + yt yi0
Computer Vision
Jan 24, 2013
14 / 44

xi + xt
yi + yt
=
=
xi0
yi0

Goal: minimize sum of squared residuals
C (xt , yt ) =
n
X
(rxi (xt )2 + ryi (yt )2 )
i=1
Computer Vision
Jan 24, 2013
14 / 44

xi + xt
yi + yt
=
=
xi0
yi0

C (xt , yt ) =
n
X
(rxi (xt )2 + ryi (yt )2 )
i=1
The solution is called the least squares solution
Computer Vision
Jan 24, 2013
14 / 44

xi + xt
yi + yt
=
=
xi0
yi0

C (xt , yt ) =
n
X
(rxi (xt )2 + ryi (yt )2 )
i=1

For translations, is equal to mean displacement
Computer Vision
Jan 24, 2013
14 / 44

xi + xt
yi + yt
=
=
xi0
yi0

C (xt , yt ) =
n
X
(rxi (xt )2 + ryi (yt )2 )
i=1

Computer Vision
Jan 24, 2013
14 / 44

xi + xt
yi + yt
=
=
xi0
yi0

C (xt , yt ) =
n
X
(rxi (xt )2 + ryi (yt )2 )
i=1

Computer Vision
Jan 24, 2013
14 / 44
Matrix Formulation
We can also write as a matrix equation
!!""#!#
!""#$#
!!""#$#
Solve for t by looking at the fixed-point equation
Computer Vision
Jan 24, 2013
15 / 44
When we are dealing with an affine transformation

0
a b c
x
x
y 0 = d e f y
w0
0 0 1
w
How many unknowns?
Computer Vision
Jan 24, 2013
16 / 44

0
a b c
x
x
y 0 = d e f y
w0
0 0 1
w
How many unknowns?
How many equations per match?
Computer Vision
Jan 24, 2013
16 / 44

0
a b c
x
x
y 0 = d e f y
w0
0 0 1
w
How many unknowns?
How many matches do we need?
Computer Vision
Jan 24, 2013
16 / 44

0
a b c
x
x
y 0 = d e f y
w0
0 0 1
w
How many unknowns?
Why to use more?
Computer Vision
Jan 24, 2013
16 / 44

0
a b c
x
x
y 0 = d e f y
w0
0 0 1
w
How many unknowns?
Why to use more?
Computer Vision
Jan 24, 2013
16 / 44
Affine Transformation Cost Function
We can write the residuals as

rxi (a, b, c, d, e, f ) =
ryi (a, b, c, d, e, f ) =
(axi + byi + c) xi0

(dxi + eyi + f ) yi0
Cost function
C (a, b, c, d, e, f ) =
N
X
rxi (a, b, c, d, e, f )2 + ryi (a, b, c, d, e, f )2
i=1
Computer Vision
Jan 24, 2013
17 / 44

rxi (a, b, c, d, e, f ) =
ryi (a, b, c, d, e, f ) =
(axi + byi + c) xi0

Cost function
C (a, b, c, d, e, f ) =
N
X
i=1
And in matrix form ...

Computer Vision
Jan 24, 2013
17 / 44

rxi (a, b, c, d, e, f ) =
ryi (a, b, c, d, e, f ) =
(axi + byi + c) xi0

Cost function
C (a, b, c, d, e, f ) =
N
X
i=1
And in matrix form ...

Computer Vision
Jan 24, 2013
17 / 44
Matrix form
$""#%#
!!""#$#
!!""#%#
Computer Vision
Jan 24, 2013
18 / 44
General Formulation
Let x 0 = f (x; p) be a parametric transformation

In the case of translation, similarity and affine, there is a linear relationship
between the amount of motion x = x 0 x and the unknown parameters
x = x 0 x = J(x)p
f
is the Jacobian of the transformation f with respect to the
with J = p
motion parameters p
Computer Vision
Jan 24, 2013
19 / 44
General Formulation
Let x 0 = f (x; p) be a parametric transformation

In the case of translation, similarity and affine, there is a linear relationship
between the amount of motion x = x 0 x and the unknown parameters
x = x 0 x = J(x)p
f
is the Jacobian of the transformation f with respect to the
with J = p
motion parameters p
Computer Vision
Jan 24, 2013
19 / 44
General Formulation
Lets do a couple on the board!
Computer Vision
Jan 24, 2013
20 / 44
General Formulation
The sum of square residuals is then

X
ELLS =
||J(xi )p xi ||22
i
X
X
X
p [
JT (xi )J(xi )]p 2pT [
JT (xi )xi )] +
||xi ||2
T
Computer Vision
Jan 24, 2013
21 / 44
General Formulation

X
ELLS =
||J(xi )p xi ||22
i
X
X
X
p [
JT (xi )xi )] +
||xi ||2
pT Ap 2pT b + c
Computer Vision
Jan 24, 2013
21 / 44
General Formulation

X
ELLS =
||J(xi )p xi ||22
i
X
X
X
p [
JT (xi )xi )] +
||xi ||2
pT Ap 2pT b + c
We can compute the solution by looking for a fixed point, yielding

Ap = b
with A =
JT (xi )J(xi ) the Hessian and b =
Computer Vision
JT (xi )xi
Jan 24, 2013
21 / 44
General Formulation

X
ELLS =
||J(xi )p xi ||22
i
X
X
X
p [
JT (xi )xi )] +
||xi ||2
pT Ap 2pT b + c
We can compute the solution by looking for a fixed point, yielding

Ap = b
with A =
JT (xi )J(xi ) the Hessian and b =
Computer Vision
JT (xi )xi
Jan 24, 2013
21 / 44
Uncertainty Weighting
The above solution assumes that all feature points are matched with same
accuracy.
If we associate a scalar variance i2 with each correspondence, we can
minimize the weighted least squares problem
X
i2 ||ri ||22
EWLS =
i
Computer Vision
Jan 24, 2013
22 / 44
accuracy.
X
i2 ||ri ||22
EWLS =
i
If the i2 are fixed, then the solution is simply

p = (T AT A)1 T Ab
with , the matrix containing for each observation the noise level
Computer Vision
Jan 24, 2013
22 / 44
accuracy.
X
i2 ||ri ||22
EWLS =
i

p = (T AT A)1 T Ab
What if we dont know ?
Computer Vision
Jan 24, 2013
22 / 44
accuracy.
X
i2 ||ri ||22
EWLS =
i

p = (T AT A)1 T Ab
Solve using iteratively reweighted least squares (IRLS)
Computer Vision
Jan 24, 2013
22 / 44
accuracy.
X
i2 ||ri ||22
EWLS =
i

p = (T AT A)1 T Ab
Solve using iteratively reweighted least squares (IRLS)
Computer Vision
Jan 24, 2013
22 / 44
Homographies
!#"
!"
To unwarp (rectify) and image

solve for homography H given p and p 0
Computer Vision
Jan 24, 2013
23 / 44
Homographies
!#"
!"

solve equations of the form: wp0 = Hp
Computer Vision
Jan 24, 2013
23 / 44
Homographies
!#"
!"

linear in unknowns: w and coefficients of H
Computer Vision
Jan 24, 2013
23 / 44
Homographies
!#"
!"

H is defined up to an arbitrary scale factor
Computer Vision
Jan 24, 2013
23 / 44
Homographies
!#"
!"

how many points are necessary to solve for H?
Computer Vision
Jan 24, 2013
23 / 44
Homographies
!#"
!"

how many points are necessary to solve for H?
Computer Vision
Jan 24, 2013
23 / 44
Solving for Homographies

0
axi
h00
ayi0 = h10
a
h20
h01
h11
h21

h02
xi
h12 yi
h22
1
To get to non-homogenous coordinates
xi0
yi0
h00 xi
h20 xi
h10 xi
h20 xi
+ h01 yi
+ h21 yi
+ h11 yi
+ h21 yi
Computer Vision
+ h02
+ h22
+ h12
+ h22
Jan 24, 2013
24 / 44

0
axi
h00
ayi0 = h10
a
h20
h01
h11
h21

h02
xi
h12 yi
h22
1

xi0
yi0
h00 xi
h20 xi
h10 xi
h20 xi
+ h01 yi
+ h21 yi
+ h11 yi
+ h21 yi
+ h02
+ h22
+ h12
+ h22
Warning: This is non-linear!!!
Computer Vision
Jan 24, 2013
24 / 44

0
axi
h00
ayi0 = h10
a
h20

h02
xi
h12 yi
h22
1
h01
h11
h21

xi0
yi0
h00 xi
h20 xi
h10 xi
h20 xi
+ h01 yi
+ h21 yi
+ h11 yi
+ h21 yi
+ h02
+ h22
+ h12
+ h22

But wait a minute!
xi0 (h20 xi + h21 yi + h22 )
= h00 xi + h01 yi + h02
yi0
= h10 xi + h11 yi + h12
(h20 xi + h21 yi + h22 )
Computer Vision
Jan 24, 2013
24 / 44

0
axi
h00
ayi0 = h10
a
h20

h02
xi
h12 yi
h22
1
h01
h11
h21

xi0
yi0
h00 xi
h20 xi
h10 xi
h20 xi
+ h01 yi
+ h21 yi
+ h11 yi
+ h21 yi
+ h02
+ h22
+ h12
+ h22

But wait a minute!
xi0 (h20 xi + h21 yi + h22 )
= h00 xi + h01 yi + h02
yi0
= h10 xi + h11 yi + h12
(h20 xi + h21 yi + h22 )
Computer Vision
Jan 24, 2013
24 / 44
Solving for homographies

xi0 (h20 xi + h21 yi + h22 )
= h00 xi + h01 yi + h02
yi0
= h10 xi + h11 yi + h12
(h20 xi + h21 yi + h22 )
This is still linear in the unknowns
Computer Vision
Jan 24, 2013
25 / 44

Taking all the observations into account
!"#$#%#
%#
!"#
Defines a least squares problem:

min ||Ah||22
h
Computer Vision
Jan 24, 2013
26 / 44

!"#$#%#
%#
!"#

min ||Ah||22
h
Since h is only defined up to scale, solve for unit vector
Computer Vision
Jan 24, 2013
26 / 44

!"#$#%#
%#
!"#

min ||Ah||22
h

= eigenvector of AT A with smallest eigenvalue
Solution: h
Computer Vision
Jan 24, 2013
26 / 44

!"#$#%#
%#
!"#

min ||Ah||22
h

Solution: h
Works with 4 or more points
Computer Vision
Jan 24, 2013
26 / 44

!"#$#%#
%#
!"#

min ||Ah||22
h

Solution: h
Works with 4 or more points
Computer Vision
Jan 24, 2013
26 / 44
Image Alignment Algorithm
Given images A and B

1
Compute image features for A and B
Match features between A and B
Compute homography between A and B using least squares on set of

matches
Is there a problem with this?
Computer Vision
Jan 24, 2013
27 / 44
Image Alignment Algorithm
Given images A and B

1
Compute image features for A and B
Match features between A and B
Compute homography between A and B using least squares on set of

matches
Is there a problem with this?
Computer Vision
Jan 24, 2013
27 / 44
Robustness
!"#$%&'()
%*$%&'()
Computer Vision
Jan 24, 2013
28 / 44
Simple case
Lets consider a simpler example ... linear regression
!"#$%&'()*+,)-)%+.&),#),/&0&)1-,-2#+.,0)
3&-0,)045-"&0)6,)
How can we fix this?
Computer Vision
Jan 24, 2013
29 / 44
Simple case
!"#$%&'()*+,)-)%+.&),#),/&0&)1-,-2#+.,0)
3&-0,)045-"&0)6,)

We need a better cost function
Computer Vision
Jan 24, 2013
29 / 44
Simple case
!"#$%&'()*+,)-)%+.&),#),/&0&)1-,-2#+.,0)
3&-0,)045-"&0)6,)

We need a better cost function
Computer Vision
Jan 24, 2013
29 / 44
More Robust Least-squares

Least-squares assumes that the noise follows a Gaussian distribution
M-estimators are use to make least-squares more robust
Computer Vision
Jan 24, 2013
30 / 44

They involve applying a robust penalty function (r) to the residuals
X
(||ri ||)
ERLS (p) =
i
instead of taking the square of the residual
Computer Vision
Jan 24, 2013
30 / 44

X
(||ri ||)
ERLS (p) =
i

We can take the derivative with respect to p and set it to 0
X
||ri || X (||ri ||) T ri
=
r
=0
(||ri ||)
p
||ri || i p
i
where (r) = 0 (r) is the derivative, called influence function
Computer Vision
Jan 24, 2013
30 / 44

X
(||ri ||)
ERLS (p) =
i

X
||ri || X (||ri ||) T ri
=
r
=0
(||ri ||)
p
||ri || i p
i

If we introduce a weight w (r ) = (r )/r , we observe that finding the
stationary point is equivalent to minimizing the iteratively reweighted
least squares (IRLS)
X
EIRLS =
w (||ri ||)||ri ||2
i
Computer Vision
Jan 24, 2013
30 / 44

X
(||ri ||)
ERLS (p) =
i

X
||ri || X (||ri ||) T ri
=
r
=0
(||ri ||)
p
||ri || i p
i

If we introduce a weight w (r ) = (r )/r , we observe that finding the
stationary point is equivalent to minimizing the iteratively reweighted
least squares (IRLS)
X
EIRLS =
w (||ri ||)||ri ||2
i
Computer Vision
Jan 24, 2013
30 / 44
Iterative reweighted least-squares
We want to minimize
EIRLS =
w (||ri ||)||ri ||2
A simple algorithm works by iterating between

1
2
Solving for the parameters p

Solving for the weights w
Computer Vision
Jan 24, 2013
31 / 44
We want to minimize
EIRLS =
w (||ri ||)||ri ||2

1
2

When the number of outliers is very high, IRLS does not work well (will not
converge to the global optima)
Computer Vision
Jan 24, 2013
31 / 44
We want to minimize
EIRLS =
w (||ri ||)||ri ||2

1
2

When the number of outliers is very high, IRLS does not work well (will not
converge to the global optima)
Computer Vision
Jan 24, 2013
31 / 44
Simple Idea
Given a hypothesized line, count the number of points that agree with the
line
Agree = within a small distance of the line i.e., the inliers to that line
Computer Vision
Jan 24, 2013
32 / 44
Simple Idea
line
For all possible lines, select the one with the largest number of inliers
Computer Vision
Jan 24, 2013
32 / 44
Simple Idea
line
For all possible lines, select the one with the largest number of inliers
Computer Vision
Jan 24, 2013
32 / 44
Counting Inliers
!"#$%&'()*)
Computer Vision
Jan 24, 2013
33 / 44
Counting Inliers
!"#$%&'()*+)
Computer Vision
Jan 24, 2013
33 / 44
Counting Inliers
!"#$%&'()*+)
Whats the problem with this approach?
Computer Vision
Jan 24, 2013
33 / 44
How do we find the best line?
Unlike least-squares, no simple closed-form solution

Hypothesize-and-test
Computer Vision
Jan 24, 2013
34 / 44

Try out many lines, keep the best one
Computer Vision
Jan 24, 2013
34 / 44

Which lines?
Computer Vision
Jan 24, 2013
34 / 44

Which lines?
Computer Vision
Jan 24, 2013
34 / 44
Translations
Computer Vision
Jan 24, 2013
35 / 44
RAndom SAmple Consensus
!"#"$%&!"#&'(%$)&(%&*(+,-'.&$-/+%&$"%$#&'&
Computer Vision
Jan 24, 2013
36 / 44
!"#"$%&'()%*"+&,'%$*&'%&+'(-),.&$)/(%&!"#!$%&&
Computer Vision
Jan 24, 2013
36 / 44
!"#$"#%#&'%#()*+,)-.*%/0#&%#&'%&01&'+#%*"23'(%.4%0*,0'(+%
Computer Vision
Jan 24, 2013
36 / 44
RANSAC
All the inliers will agree with each other on the translation vector; the
(hopefully small) number of outliers will (hopefully) disagree with each other
RANSAC only has guarantees if there are < 50% outliers
Computer Vision
Jan 24, 2013
37 / 44
RANSAC
All good matches are alike; every bad match is bad in its own way.
[Tolstoy via Alyosha Efros]
Computer Vision
Jan 24, 2013
37 / 44
RANSAC
All good matches are alike; every bad match is bad in its own way.
[Tolstoy via Alyosha Efros]
Computer Vision
Jan 24, 2013
37 / 44
RANSAC for line fitting example

1
Randomly select minimal

subset of points
Hypothesize a model
[Source: R. Raguram]
Computer Vision
Jan 24, 2013
38 / 44

1

subset of points
Hypothesize a model
Compute error function
Computer Vision
Jan 24, 2013
38 / 44

1

subset of points
Hypothesize a model
Select points consistent

with model
Computer Vision
Jan 24, 2013
38 / 44

1

subset of points
Hypothesize a model

with model
Repeat hypothesize and

verify loop
Computer Vision
Jan 24, 2013
38 / 44

1

subset of points
Hypothesize a model

with model

verify loop
Computer Vision
Jan 24, 2013
38 / 44

1

subset of points
Hypothesize a model

with model

verify loop
Choose model with

largest set of inliers
Computer Vision
Jan 24, 2013
38 / 44

1

subset of points
Hypothesize a model

with model

verify loop
Choose model with

largest set of inliers
Computer Vision
Jan 24, 2013
38 / 44
RANSAC
Inlier threshold related to the amount of noise we expect in inliers

Often model noise as Gaussian with some standard deviation (e.g., 3 pixels)
Computer Vision
Jan 24, 2013
39 / 44
RANSAC

Number of rounds related to the percentage of outliers we expect, and the
probability of success wed like to guarantee
Computer Vision
Jan 24, 2013
39 / 44
RANSAC

Suppose there are 20% outliers, and we want to find the correct answer with
99% probability
Computer Vision
Jan 24, 2013
39 / 44
RANSAC

99% probability
How many rounds do we need?
Computer Vision
Jan 24, 2013
39 / 44
RANSAC

99% probability
How many rounds do we need?
Computer Vision
Jan 24, 2013
39 / 44
How many rounds?

Sufficient number of trials S must be tried.
Let p be the probability that any given correspondence is valid and P be the
total probability of success after S trials.
Computer Vision
Jan 24, 2013
40 / 44
How many rounds?

The likelihood in one trial that all k random samples are inliers is p k
Computer Vision
Jan 24, 2013
40 / 44
How many rounds?

The likelihood that S such trials will all fail is
1 P = (1 p k )S
Computer Vision
Jan 24, 2013
40 / 44
How many rounds?

1 P = (1 p k )S
The required minimum number of trials is

S=
log(1 P)
log(1 p k )
Computer Vision
Jan 24, 2013
40 / 44
How many rounds?

1 P = (1 p k )S

S=
log(1 P)
log(1 p k )
The number of trials grows quickly with the number of sample points used.
Computer Vision
Jan 24, 2013
40 / 44
How many rounds?

1 P = (1 p k )S

S=
log(1 P)
log(1 p k )
Use the minimum number of sample points k possible for any given trial
Computer Vision
Jan 24, 2013
40 / 44
How many rounds?

1 P = (1 p k )S

S=
log(1 P)
log(1 p k )
Use the minimum number of sample points k possible for any given trial
Computer Vision
Jan 24, 2013
40 / 44
How big is the number of samples?

For alignment, depends on the motion model
Each sample is a correspondence (pair of matching points)
Computer Vision
Jan 24, 2013
41 / 44
RANSAC pros and cons

Pros
Simple and general
Applicable to many different problems
Computer Vision
Jan 24, 2013
42 / 44

Pros
Simple and general
Often works well in practice
Computer Vision
Jan 24, 2013
42 / 44

Pros
Simple and general
Cons
Parameters to tune
Computer Vision
Jan 24, 2013
42 / 44

Pros
Simple and general
Cons
Parameters to tune
Sometimes too many iterations are required
Computer Vision
Jan 24, 2013
42 / 44

Pros
Simple and general
Cons
Parameters to tune
Can fail for extremely low inlier ratios
Computer Vision
Jan 24, 2013
42 / 44

Pros
Simple and general
Cons
Parameters to tune
We can often do better than brute-force sampling
Computer Vision
Jan 24, 2013
42 / 44

Pros
Simple and general
Cons
Parameters to tune
We can often do better than brute-force sampling
Computer Vision
Jan 24, 2013
42 / 44
RANSAC as Voting
An example of a voting-based fitting scheme

Each hypothesis gets voted on by each data point, best hypothesis wins
Computer Vision
Jan 24, 2013
43 / 44
RANSAC as Voting

There are many other types of voting schemes, e.g., Hough transforms
Computer Vision
Jan 24, 2013
43 / 44
RANSAC as Voting

There are many other types of voting schemes, e.g., Hough transforms
Computer Vision
Jan 24, 2013
43 / 44
Next class ... more on cameras and projection
Computer Vision
Jan 24, 2013
44 / 44

Lecture 06

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Lecture 06

Uploaded by

Copyright:

Available Formats

Computer Vision: Image Alignment

Jan 24, 2013

Raquel Urtasun (TTI-C)

Jan 24, 2013

Chapter 2.1, 3.6, 4.3 and 6.1 of Szeliskis book

Raquel Urtasun (TTI-C)

Jan 24, 2013

What did we see in class last week?

Raquel Urtasun (TTI-C)

Jan 24, 2013

What is the geometric relationship between these images?

Raquel Urtasun (TTI-C)

Jan 24, 2013

What is the geometric relationship between these images?

Very important for creating mosaics!

Raquel Urtasun (TTI-C)

Jan 24, 2013

Image warping: change domain of image

Jan 24, 2013

Parametric (global) warping

Transformation T is a coordinate-changing machine:

Jan 24, 2013

Forward and Inverse Warping

Raquel Urtasun (TTI-C)

Jan 24, 2013

All 2D Linear Transformations

Linear transformations are combinations of

Raquel Urtasun (TTI-C)

Jan 24, 2013

All 2D Linear Transformations

Raquel Urtasun (TTI-C)

Jan 24, 2013

All 2D Linear Transformations

Raquel Urtasun (TTI-C)

Jan 24, 2013

All 2D Linear Transformations

Raquel Urtasun (TTI-C)

Jan 24, 2013

All 2D Linear Transformations

Raquel Urtasun (TTI-C)

Jan 24, 2013

All 2D Linear Transformations

What about the translation?

Raquel Urtasun (TTI-C)

Jan 24, 2013

All 2D Linear Transformations

What about the translation?

Raquel Urtasun (TTI-C)

Jan 24, 2013

Properties of affine transformations:

Jan 24, 2013

Properties of affine transformations:

Jan 24, 2013

These transformations are a nested set of groups

Jan 24, 2013

Jan 24, 2013

Least squares formulation

We define the residuals as

Raquel Urtasun (TTI-C)

Jan 24, 2013

Least squares formulation

We define the residuals as

(rxi (xt )2 + ryi (yt )2 )

Raquel Urtasun (TTI-C)