More Single View Geometry 1
2008-1
Multi View Geometry (Spring '08)
Prof. Kyoung Mu Lee
SoEECS, Seoul National University
More Single View Geometry
More Single View Geometry 2
Actions of a Projective Camera on Planes
• Planar homography 9 For a point on
• The most general transformation that can occur between a scene plane and an image plane under perspective imaging is a plane projective transformation
(affine camera-affine transformation)
Hxπ
=
planar homography
π
More Single View Geometry 3
Multi View Geometry (Spring '08) K. M. Lee, EECS, SNU
Actions of a Projective Camera on Lines
• Forward projection: Line to line
• Back-projection: line to plane
l P l→ T
l l
P l P X
l x PX x
line the of plane projected -
back the is
0 ) (
0 and
T T T
T
⇒
=
⇒
= l =
PT π=
More Single View Geometry 4
Actions of a Projective Camera on Conics
• Back projection: conic to cone
• Ex:
CP P Q C→ CO= T
0 PC C P C Q
Q C
Q
=
= ( )
, of vector null the is and 3 rank has
T CO
CO CO
C CP
P
X CP P X
Cx x PX x
conic the of cone projected -
back the is
0 ) (
0
and
T T T
T
⇒
=
⇒
=
=
[ ] [ ] ⎥
⎦
⎢ ⎤
⎣
=⎡
⎥⎦
⎢ ⎤
⎣
=⎡
= | , | 0
T T
co 0
0 CK 0 K
K 0 C Q K 0 I K P
More Single View Geometry 5
Multi View Geometry (Spring '08) K. M. Lee, EECS, SNU
Images of Smooth Surfaces
• The contour generator Γ is the set of points X on S at which rays are tangent to the surface. The corresponding apparent contour γ is the set of points x which are the image of X, i.e. γ is the image of Γ
• The contour generator Γ depends only on position of projection center, γ depends also on rest of P
More Single View Geometry 6
Actions of a Projective Camera on Quadrics
• Forward Projection: Quadric to conic
• The cone with vertex V and tangent to the quadric Q (degenerate quadric) is
QC πΓ=
−
ip relationsh polar
pole
0 V QCO =
T T
CO (V QV)Q-(QV)(QV) Q =
More Single View Geometry 7
Multi View Geometry (Spring '08) K. M. Lee, EECS, SNU
Fixed Camera Center
• Projective imaging devices
~]
| [' ' '
~],
|
[I C P KR I C
KR
P= − = −
homography planar
,
: = ′ ′ −1
′=
⇒x Hx H (KR)(KR)
( )KR P R
K
P'= ' ' -1
⇒
More Single View Geometry 8
Moving the image plane (zooming)
• Increase in focal length:
9 move image plane along the principal axis 9 magnification
f k f′
=
More Single View Geometry 9
Multi View Geometry (Spring '08) K. M. Lee, EECS, SNU
Camera rotation
x KRK 0]X
| K[I KRK 0]X
| K[R x'
X 0 I K x
-1
-1 =
=
=
= [ | ]
rotation conjugate
1 : KRK H= −
{
eiθ e−iθ}
=
−values μ,μ ,μ three
angle rotation the
is
point), (vanishing axis
rotation the
of direction the
represent
1
θ e
More Single View Geometry 10
• Compute the homography that warps the image quadrilateral to a rectangle with the correct aspect ratio
• warp the image
Applications – Synthetic Views
More Single View Geometry 11
Multi View Geometry (Spring '08) K. M. Lee, EECS, SNU
Applications – Planar panoramic mosaicing
More Single View Geometry 12
• close-up: interlacing problem
Applications – Planar panoramic mosaicing
More Single View Geometry 13
Multi View Geometry (Spring '08) K. M. Lee, EECS, SNU
Applications – Planar panoramic mosaicing
More Single View Geometry 14
Projective (reduced) notation
• If canonical coordinates are chosen
• Then the camera matrix becomes
and the camera center is
z This implies that all images with the same camera center are projectively equivalent
T 4
T 3
T 2
T
1 (1,0,0,0) ,X (0,1,0,0) ,X (0,0,1,0) ,X (0,0,0,1)
X = = = =
T 4
T 3
T 2
T
1 (1,0,0) ,x (0,1,0) ,x (0,0,1) ,x (1,1,1)
x = = = =
⎥⎥
⎦
⎤
⎢⎢
⎣
⎡
−−
= −
d
c d
b d
a 0
0 0
0 0 0 P
d T
c b
a , , , ) (
C= −1 −1 −1 −1
More Single View Geometry 15
Multi View Geometry (Spring '08) K. M. Lee, EECS, SNU
Moving the Camera Center
• When the camera center moves –motion parallaxoccurs
epipolar line
More Single View Geometry 16
What does calibration give?
•
• The angle between two rays
• A calibrated camera (K is known) is a direction sensor which can measure the direction of rays
• An image line l defines a plane through the camera center with normal n=KTl measured in the camera’s Euclidean frame
( )( )
( )(
2)
1 - T T - 2 1 1 - T T - 1
2 1 - T T - 1
2 T 2 1 T 1
2 T 1
) ( ) (
) ( cos
x K K x x K K x
x K K x
d d d d
d d
= θ=
x K d
Kd d
0
| I K x
x d
X
1
T T,1) ](
[
then , of points projected back
the be
~ Let
= −
⇒
=
=
= λ λ
in the camera Euclidean coordinates
) 0
) ( 0 ) ( 0
(dTn= ⇒ K−1xTn= ⇒xT K−Tn = ⇒l=K−Tn
More Single View Geometry 17
Multi View Geometry (Spring '08) K. M. Lee, EECS, SNU
The image of the Absolute Conic
• mapping between π∞to an image is given by the planar homogaphy x=Hd, with H=KR
• image of the absolute conic (IAC) is
Independent on C
1 1
1
) ( ) ( ) (
on conic,
absolute an
For
−
−
−
∞
∞
−
−
=
=
=
=
′=
T T
T
KK KR
I KR ω
π I Ω C CH
H C
(
T)
1 -T -1ω= KK − =K K
More Single View Geometry 18
The image of the Absolute Conic
• Remarks:
9 depends only on the internal parameters K 9 The angle between two rays
9 Dual (line) conic (DIAC):
9 Once is identified K can be determined uniquely by Cholesky factorization
9 A plane intersects in a line, and this line intersects in two points (circular points of ). The imaged circular points lie on at the points at which the vanishing line of the plane intersects .
unchanged under any projective transformation of the image
Image of Q*∞
ω
or *
ω ω
π π∞
π ∞
Ω
π ω
ω
More Single View Geometry 19
Multi View Geometry (Spring '08) K. M. Lee, EECS, SNU
A simple calibration device
• Calibration steps:
More Single View Geometry 20
Orthogonality in the image
• Using
2 0
1ωx = xT 2 0
1ωx =
xT l=ωx
( )( )
⇒=
2 T 2 1 T 1
2 T
cos 1
ωx x ωx x
ωx θ x
1
and
= − ∞ −
=Hd ω H Ω H
xi i T
More Single View Geometry 21
Multi View Geometry (Spring '08) K. M. Lee, EECS, SNU
Vanishing Points
• Let the point on a line through A with direction be
• Then
• And
T T,0) (d D= D
A
X(λ)= +λ P=K
[ ]
I|0More Single View Geometry 22
Vanishing Points
• Camera rotation from vanishing points
• Assume two cameras with same K and rotation by R
• Vanishing points are depend on camera rotation but not position
• Thus, two corresponding directions are sufficient for determining R
i i
i i i
i i
i i i
i i
d R d
v K
v d K
d K v
v K
v d K
Kd v
ˆ ˆ
ˆ ˆ
1 1 1 1
′=
′
= ′
⇒ ′
= ′
′
=
⇒
=
−
−
−
−
More Single View Geometry 23
Multi View Geometry (Spring '08) K. M. Lee, EECS, SNU
Vanishing Lines
• All planes with normal n have vanishing line
• Or
n K l= −T l
K n= T
More Single View Geometry 24
Vanishing Lines
More Single View Geometry 25
Multi View Geometry (Spring '08) K. M. Lee, EECS, SNU
Orthogonality relations
• The angle between two scene lines using vanishing points:
• The angle between two scene planes using vanishing lines:
( )(
2)
T 2 1 T 1
2 T
cos 1
ωv v ωv v
ωv
= v θ
( )(
2)
T 2 1 T 1
2 T
cos 1
l ω l l ω l
l ω l
∗
∗
= ∗
θ
More Single View Geometry 26
Orthogonality relations
• Orthogonality relations:
9 The vanishing points of two perpendicular lines
9 The vanishing point of the normal direction to a plane and the plane vanishing line
9 The vanishing lines of two perpendicular planes
2 0
1ωv = vT
0 ωv l ωv l l ω
v= ∗ ⇒ = ⇒ ×( )=
2 0
1ω∗l = lT
More Single View Geometry 27
Multi View Geometry (Spring '08) K. M. Lee, EECS, SNU
Affine 3D measurements and Reconstruction
More Single View Geometry 28
Affine 3D measurements and Reconstruction
More Single View Geometry 29
Multi View Geometry (Spring '08) K. M. Lee, EECS, SNU
Affine 3D measurements and Reconstruction
• Measuring a person’s height in a single image
More Single View Geometry 30
Determining K from a single view
• Using five pairs of of vanishing points of perpendicular lines
• Similarly to pairs of vanishing point & lines, and vanishing lines & lines
•
• Thus, for a camera with zero skew and square pixels, three orthogonal vanishing points are sufficient for determining K
=0
j T iωv
v ω=K−TK−1 K
[ ]
21 12 22
11
21 12 1
12
then , addition
in If
. 0 then
and 0 If
ω ω α
α
ω ω ω
=
=
=
=
=
=
=
=
=
= − −
y x
T ij
K K K
s ω K K
More Single View Geometry 31
Multi View Geometry (Spring '08) K. M. Lee, EECS, SNU
Determining K from a single view
• Three sources of constraints on
9 Metric plane imaged with known homography
9 Vanishing points and lines corresponding to perpendicular directions
9 “internal constraints” such as zero skew or square pixels ω
More Single View Geometry 32
Determining K from a single view
More Single View Geometry 33
Multi View Geometry (Spring '08) K. M. Lee, EECS, SNU
Determining K using vanishing points and lines
More Single View Geometry 34
Determining K using vanishing points and lines
v1 v1
v1
v1
v2
More Single View Geometry 35
Multi View Geometry (Spring '08) K. M. Lee, EECS, SNU
Determining K using vanishing points and lines
More Single View Geometry 36
Determining K using vanishing points and lines
•
• Only one constraints is needed to determine f (and thus K)
• We can synthesize a fronto-parallel view using K and vanishing line l
⎟⎟⎠
⎜⎜ ⎞
⎝
= ⎛
=
=
=
= 1 1 1
then , 0 and
, ,
0 Assume
2 2
0 0
f diag f
y x
s x y
ω
α α
More Single View Geometry 37
Multi View Geometry (Spring '08) K. M. Lee, EECS, SNU
Single View Reconstruction
• Piecewise planar graphical model from a single image
9 Assume three orthogonal planes
9 Determine K using three vanishing points and square pixels 9 Proportioned width and height are determined by rectification 9 Find homography for texture mapping for each plane using K and l
More Single View Geometry 38
Calibrating Conic
• Image of a cone with apex angle 45oand axis coinciding with the principal axis
• Let and the point on the cone be , then the calibrating conic is an affine transformation of a unit circle
[ ]I 0 K
P= | X2+Y2 =Z2
Unit circle
More Single View Geometry 39
Multi View Geometry (Spring '08) K. M. Lee, EECS, SNU
Calibrating Conic
• Orthogonality:
Thus, x and x’ are orthogonal if
x C x CSx x
x CS x ωx x
&
T T
T T
= ′
= ′
= ′
= ′ −1
0
] 1 , 1 , 1 [
1, = −
=K−TDK− D diag C
C x Sx
x
of center the w.r.t.
of point reflection
= ;
&
More Single View Geometry 40
Determining K using vanishing points and lines
• Finding calibrating conic using orthogonal vanishing points