Upload
roxy
View
26
Download
0
Embed Size (px)
DESCRIPTION
Lecture 11. Adding Edge Element to Constraint Coarse-to-Fine Approach Optical Flow. Edge Constraint for Noise Removal. Vertical Edge Element ON/OFF. Horizontal Edge Element ON/OFF. Tends to create problems at edges If edge element “ON”, turn off smoothness. - PowerPoint PPT Presentation
Citation preview
Lecture 11Lecture 11
Adding Edge Element to Constraint
Coarse-to-Fine Approach
Optical Flow
2
Edge Constraint for Noise RemovalEdge Constraint for Noise Removal
ijijijijijjim n
ijij vIIhIIIFE
11 211
211
1
1
1
1
2
Vertical Edge ElementON/OFF
Horizontal Edge ElementON/OFF
Tends to create problems at edgesIf edge element “ON”, turn off smoothness
""322 esShapeOfEdgvh ijij
3
Edge Constraint for Noise RemovalEdge Constraint for Noise Removal
Cost3 = 0
Cost3 = 1
Cost3 = 0.1
4
Edge Shape PrecedencesEdge Shape Precedences
Vij
HijDij
Dij Hij Vij
5
AlgorithmAlgorithm
1. Start Temperature T is high2. Initialize D(i,j) = Random 0..103. For each pixel (i,j)
if Dij = 0, Vij = 0, Hij = 0; E0= ... ; P0 = e - E0 /Tif Dij = 0, Vij = 0, Hij = 1; E1 = ... ; P1 = e - E1 /T
.
.if Dij = 10, Vij = 1, Hij = 1; E44 = ... ; P44 = e – E44 /TFor each Pi = Pi/Z Pi/sum(Pi)
Sample for state K from pdf Pi Dij = 1, Vij = 0, Hij = 0
4. Reduce T5. Repeat step 3-4 while E is not stable
6
Coarse-to-FineCoarse-to-Fine
Coarse
FineDm = 20
Dm1/2 = 10
Dm1/4 = 5Solve First
Left Right D
Dhalf
D1/4
101
41
2
7
Coarse-to-Fine - PyramidCoarse-to-Fine - Pyramid
8
Stereo – Parallel AxisStereo – Parallel Axis
Matching Points will be in the same row
Same image Plane
One Point RijX
9
Stereo – Parallel AxisStereo – Parallel Axis
Epipolar – Line formed by intersection of epipolar plane with image plane
Epipolar Plane – Planes defined by 2 camera center and viewed 3-D point
10
Motion – Optical FlowMotion – Optical Flow
Movement of Image Points
Optical Flow is written as (u,v)tt xxx
dt
dxu 1
tt yyydt
dyv 1
Frame 1 Frame 2
11
Motion Correspondence ProblemMotion Correspondence Problem
1. Start Temperature T is high2. u = Random (-5,-4,….,0,…., 4, 5); v = Random (-5,-4,….,0,…., 4, 5) 3. For every (edge) pixel
if u = -5, v = -5; E0= ... ; P0 = e - E0 /Tif u = -5, v = -4; E1 = ... ; P1 = e - E1 /T
.
.if u = 5, v = 5; E121 = ... ; P121 = e – E121 /TFor each Pi = Pi/Z Pi/sum(Pi)
Sample for (u,v) from pdf Pi4. Reduce T5. Repeat step 3-4 while E is not stable
12
Stereo Matching : DisparityStereo Matching : Disparity
2),()1,(12
),(),1(1
1
1
1
1
21),(
2),( yxyxyxyx
m nvyuxyx uuuuIIE
2),()1,(12
),(),1(1 yxyxyxyx vvvv
Can add 1-1 match constraint
13
Barber Pole IllusionBarber Pole Illusion
Aperture Problem – Perceive that motion is perpendicular to edge