Computer vision

3D Transformations

Le Thanh Ha, Ph.D

Assoc. Prof. at University of Engineering and Technology,
Vietnam National University
ltha@vnu.edu.vn; lthavnu@gmail.com; 0983 692 592
 Coordinate Frames
Y0

X0

2/23/2023
 Coordinate Frames
• Interested in Some Point described by a Vector R
• We need to express R which we know in Frame 1 in Frame 0

x1 = R  sin( )
y1 = R  cos( ) Y0

x0 = x1  cos(1 ) − y1  sin(1 ) R

y0 = x1  sin(1 ) + y1  cos(1 ) 

Equation is independent of values

−1
X0

2/23/2023
 Coordinate Frames
• Now lets put it in matrix form
x0 = x1  cos(1 ) − y1  sin(1 )
y0 = x1  sin(1 ) + y1  cos(1 )

 x0  cos(1 ) − sin(1 )  x1 
 y  =  sin( ) cos( )    y 
 0  1 1   1

 x0   x1 
 y  = T   y 
 0  1

• So what if we want to map the other way?

 x1  −1  x0 
 y  = T    y 
 1  0

• What is the inverse of T? Why?

2/23/2023
 Coordinate Frames
• If we look at the columns and rows of T we see that they have a norm of
one.
• Also if we take the dot product of the columns we find they are orthogonal
to each other.
• So T is an ortho-normal Matrices. Thus its transpose is its inverse.
cos(1 ) − sin(1 )  cos(1 ) sin(1 ) 
T  =   T 
−1
= 
 sin(1 ) cos(1 )  − sin(1 ) cos(1 )

• This was a simple 2DOF example what about 3.

• If we project a Z axes out the plane generated by the X and Y axes, then a
rotation around the Z axes will not affect the Z position of the vector R.

2/23/2023
 Coordinate Frames
• The 3D transformation axes about the Z axes is:
cos( ) − sin( ) 0
T z =  sin( ) cos( ) 0
 0 0 1

• Similarly for rotations around the X or Y axes we get

1 0 0   cos( ) 0 sin(  ) 
T x = 0 cos( ) − sin( ) T Y =  0 1 0 
0 sin( ) cos( )  − sin(  ) 0 cos( )

• Shorthand notation is often used

cos(x)=c x sin(x)=s x

2/23/2023
 Coordinate Transformations
• For multiple transformation we simply multiply by more matrices.
• If we have multiple rotations about the same axes we can just add the angles of the matrices. Does this make sense.

cos( 𝜑) − sin( 𝜑) 0 cos( 𝛾) − sin( 𝛾) 0

𝑇(𝛾 + 𝜑) 𝑧 = sin( 𝜑) cos( 𝜑) 0 ⋅ sin( 𝛾) cos( 𝛾) 0
0 0 1 0 0 1

cos( 𝜑) ⋅ cos( 𝛾) − sin( 𝜑) ⋅ sin( 𝛾) − cos( 𝜑) ⋅ sin( 𝛾) − sin( 𝜑) ⋅ cos( 𝛾) 0

= sin( 𝜑) ⋅ cos( 𝛾) + cos( 𝜑) ⋅ sin( 𝛾) − sin( 𝜑) ⋅ sin( 𝛾) + cos( 𝜑) ⋅ cos( 𝛾) 0
0 0 1

sin( 𝜑 + 𝛾) = sin( 𝜑) cos( 𝛾) + cos( 𝜑) sin( 𝛾)

cos( 𝜑 + 𝛾) = cos( 𝜑) cos( 𝛾) − sin( 𝜑) sin( 𝛾)

cos( 𝜑 + 𝛾) − sin( 𝜑 + 𝛾) 0 Substituting we get

𝑇(𝛾 + 𝜑) 𝑧 = sin( 𝜑 + 𝛾) cos( 𝜑 + 𝛾) 0
0 0 1

2/23/2023
 Coordinate Transformation
• So now we can express a vector in a reference frame rotated to
an arbitrary angle in space.
• What if we want to express it in a translated frame

2/23/2023
 Translating Coordinate Frames
Y1

Y0

X1

X0

2/23/2023
 Translating Coordinate Frames
• Now we can translate and rotate. Y0
Y1

x0 = x1 + x R

y0 = y1 + y
X1

• Or in terms of vectors X0

R0 = R1 + R

2/23/2023
 What if we have a rotation and a translation
• First we can rotate the frames
Y0’
• then we can translate R


Y0

 x0   x1  x 
 y  = T   y  + y  
 0  1   1 X0’

X0

2/23/2023
 Homogeneous Transformation Matrices
C1 − S1 0
• 3x3 Rotation Matrix T1 =  S1 C1 0
 0 0 1

• 3x1 Displacement Vector

 x1 
R1 =  y1 
 z1 

• For a displacement and a rotation

R0  = T  R1 + R
2/23/2023
 4x4 Homogeneous Matrix

• If we want to perform a rotation and a translation with one

operation
 x0   x1  x  C1 − S1 0
 y  = T   y  + y  T =  S1 C1 0
 0  1  
 z 0   z1   z   0 0 1

• We can create a homogeneous Transformation Matrix

C1 − S1 0 x0   x0   x1 
 S1 C1 0 y0  y  y 
TH =    = TH   1 
0
0 0 1 z0  z0   z1 
     
0 0 0 1   
0 0
2/23/2023
 Homogeneous Transformation Matrices
1 0 0 x
0 0 y 
• What does a pure translation look like TT = 
1
0 0 1 z
 
0 0 0 1

C1 − S1 0 0
• What does a pure rotation look like  S1 C1 0 0
TR = 
0 0 1 0
 
0 0 0 1

• Multiply to move from one coordinate to another.

2/23/2023
 Notation

• Coordinate systems are represented with brackets

{B}, {0}, etc.
• Vectors
Frame of Reference
– Lets Look at a Vector P
Described in Frame A  px 
A
P =  p y 
– Leading Subscript describes the frame  p z 
in which the Vector is described or
Referenced
– Individual Elements of a vector
are described by a trailing subscript

2/23/2023
 Matrix Notation

New Frame of Reference

A
A T
B R B

Original Frame of Reference

2/23/2023
 Homogenous Transformations Represent 3 Things
• Describe a Frame
• Map from one Frame to another
• Act as an Operator to move within a Frame

A
B T

2/23/2023
 Transforms Describe Frames
C1 − S1 0 A
PBorgX 
 
S1 A
• Frames can be described by
A Homogenous Transformation
A
BT =  C1
 0 B0
R 0
1
A

A
PBorgY 
PBorgZ 
 
Matrices  0 0 0 1 

• Description of Frame
– Columns of BA R are the Unit Vectors defining the directions of the principle axes of {B}
in terms of {A}  B Xˆ A 
A
B R=  Xˆ
A
B YˆB
A A ˆ
 
Z B =  BYˆA 

 B Zˆ 
 A
A
– Rows of R are the Unit Vectors defining the directions of the principle axes of {A} in
B
terms of {B}
– A PBorg is the location of the origin of {B} in terms or {A}

2/23/2023
 Mapping Between Frames
• Maps vector from Frame {B} to Frame {A}
C1 − S1 0 A
PBorgX 
• A
will rotate a vector to project its
R 
S1 A

B
components originally described in
A
BT =  C1
 0 B0
R 0
1
A

A
PBorgY 
PBorgZ 
 
{B} in the {A} of Frame  0 0 0 1 

• A
will translate the vector to adjust its origin from frame {B} to
PBorg
its new origin in {A}
B A
P P
2/23/2023
 Acts as an Operator within a coordinate frame
• Operator will rotate and translate a vector

• R Defines the angle to rotate about

• PBorg Defines the distance to translate

C1 − S1 0 P2orgX 
 
• Usually drop subscripts T =
0
S1 C1
0
R 0
1
P2orgY 
P2 orgZ 
 
 0 0 0 1 
• T Operates on A
P1 to create A P2

2/23/2023
 Bài tập
• Ký hiệu hệ tọa độ 3 chiều của căn phòng là N
• Ký hiệu hệ tọa độ 3 chiều của 01 camera gắn trong phòng là C
• Tọa độ của camera C trong căn phòng là (3, 5, 10) mét
• Góc xoay của camera C so với căn phòng theo từng trục x, y, z
là (20, 45, 210) độ
• Hãy thiết lập ma trận chuyển đổi (homogenerous matrix) 𝑵𝑪𝑻
để chuyển đổi tọa độ của một đối tượng trong camera sang
hệ tọa độ của căn phòng.
2/23/2023