Optical Flow Methods - University of Delaware
Download
Report
Transcript Optical Flow Methods - University of Delaware
Optical Flow Methods
CISC 489/689
Spring 2009
University of Delaware
Outline
• Review of Optical Flow Constraint, LucasKanade, Horn and Schunck Methods
• Lucas-Kanade Meets Horn and Schunck
• 3D Regularization
• Techniques for solving optical flow
• Confidence Measures in Optical Flow
Optical Flow Constraint
( x, y )
( x ut , y vt )
f ( x, y , t )
f ( x ut , y vt , t t ) f ( x, y, t )
f
f
f
f ( x, y, t ) x y t
f ( x, y , t )
x
y
t
Dividingby t and takinglimit t
0
f dx f dy f
0
x dt y dt t
f xu f y v f t 0
f ( x, y, t t )
Interpretation
v
f xu f y v f t 0
Constraint Line
u
f f t
v
T
f , f
x
y
u
f
x
fy
u
ft v 0
1
Lucas-Kanade Method
f x2
fx fy
f x ft
fx fy
f y2
f y ft
f x2
J fx fy
f x ft
f x f t u
f y f t v 0
f t 2 1
fx fy
f y2
f y ft
f x ft
f y ft
f t 2
J
*
u
E LK (u , v) K * J v
1
=
K J
Lucas-Kanade Method
• Local Method, window based
• Cannot solve for optical flow everywhere
• Robust to noise
7.5
15
Figures from Lucas/Kanade Meets Horn/Schunck: Combining Local and Global Optic
Flow Methods ANDR´ES BRUHN AND JOACHIM WEICKERT, 2005
Dense optical Flow
MinimizeE (u, v) ( f xu f y v f t ) 2 ?
Lacks Smoothness
Figures from Lucas/Kanade Meets Horn/Schunck: Combining Local and Global Optic
Flow Methods ANDR´ES BRUHN AND JOACHIM WEICKERT, 2005
Horn and Schunck Method
EHS (u, v) ( f xu f y v f t ) ( u v )dxdy
2
Euler-Lagrange Equations
2
2
Horn and Schunck Method
• Global Method
• Estimates flow everywhere
• Sensitive to noise
• Oversmooths the edges
105
106
Figures from Lucas/Kanade Meets Horn/Schunck: Combining Local and Global Optic
Flow Methods ANDR´ES BRUHN AND JOACHIM WEICKERT, 2005
Why combine them?
• Need dense flow estimate
• Robust to noise
• Preserve discontinuities
Combining the two…
u
E LK (u , v) K * J v
1
EHS (u, v) ( f xu f y v f t ) 2 ( u v )dxdy
2
2
u
2
2
ECLG (u , v) K * J v ( u v )dxdy
1
Combined Local Global Method
u
2
2
ECLG (u , v) K * J v ( u v )dxdy
Euler-Lagrange Equations 1
Average
Error
Lucas&Kanade
Standard
Deviation
4.3
(density 35%)
Horn&Schunk
9.8
16.2
Combining local
and global
4.2
7.7
Table: Courtesy - Darya Frolova, Recent progress in optical flow
Preserving discontinuities
• Gaussian Window does not preserve
discontinuities
• Solutions
– Use bilateral filtering
u
2
2
Ebil (u, v) K bil * J v ( u v )dxdy
1
– Add gradient constancy
u
2
2
E grad (u, v) K * J v ( u v ) (f f t 1 ) 2 dxdy
1
Bilateral support window
Images: Courtesy, Darya Frolova, Recent progress in optical flow
Robust statistics – simple example
Find “best” representative for the set of numbers
x
xi
L2:
E
x xi
2
min
i
Influence of xi on E:
Enew Eold 2xi x
proportional to
x xi
Outliers influence the most
x mean( xi )
L1:
E
xx
i
min
i
x i → xi + ∆
Enew Eold
equal for all xi
Majority decides
x median( xi )
Slide: Courtesy - Darya Frolova, Recent progress in optical flow
Robust statistics
many ordinary people
a very rich man
wealth
Oligarchy
Votes proportional to the wealth
like in L2 norm minimization
Democracy
One vote per person
like in L1 norm minimization
Slide: Courtesy - Darya Frolova, Recent progress in optical flow
Combination of two flow constraints
min
I
warped
I
I
warped
I
video
I warped I ( x u , y v, t 1) ; I I ( x, y, t )
usual: L2
x2
robust: L1
robust regularized
x2 2
x
ε
easy to analyze and minimize
– sensitive to outliers
robust in presence of outliers
– non-smooth: hard to analyze
smooth: easy to analyze
robust in presence of outliers
[A. Bruhn, J. Weickert, 2005]
Towards ultimate motion estimation: Combining highest accuracy withSlide:
real-time
performance
Courtesy - Darya Frolova, Recent progress in optical flow
Robust statistics
3D Regularization
• ( u v ) accounted for spatial
regularization
• If velocities do not change suddenly with time,
can we regularize in time as well?
2
2
3D Regularization
u u u
3u
x y t
v v v
3v
x y t
u
2
2
ECLG 3 (u , v)
K * J v ( 3u 3v )dxdydt
X [ 0 ,T ]
1
Multiresolution estimation
run iterative estimation
warp & upsample
run iterative estimation
.
.
.
image J1
image
Gaussian pyramid of image 1
image
Image I2
21
Gaussian pyramid of image 2
Multi-resolution Lucas Kanade
Algorithm
Compute Iterative LK at highest level
•For Each Level i
•Take flow u(i-1), v(i-1) from level i-1
•Upsample the flow to create u*(i), v*(i) matrices of twice
resolution for level i.
•Multiply u*(i), v*(i) by 2
•Compute It from a block displaced by u*(i), v*(i)
•Apply LK to get u’(i), v’(i) (the correction in flow)
•Add corrections u’(i), v’(i) to obtain the flow u(i), v(i) at the ith
level, i.e., u(i)=u*(i)+u’(i), v(i)=v*(i)+v’(i)
Comparison of errors
For Yosemite sequence with clouds
Table: Courtesy - Darya Frolova, Recent progress in optical flow
Solving the system
Au f
How to solve?
Start with some initial guess
u initial
and apply some iterative method
2 components of success:
fast convergence
good initial guess
Relaxation smoothes the error
Relaxation schemes have smoothing property:
............
Only neighboring pixels
are coupled in relaxation
scheme
It may take thousands of
iterations to propagate
information to large
distance
Relaxation smoothes the error
Examples
1D case:
2D case:
Error of initial guess
Error after 5 relaxation
Error after 15 relaxations
Idea: coarser grid
initial grid – fine grid
On a coarser grid low frequencies
become higher
Hence, relaxations can be more
effective
coarse grid – we take every second point
Multigrid 2-Level V-Cycle
1. Iterate ⇒ error becomes smooth
2. Transfer error equation to the coarse
level ⇒ low frequencies become high
5. Correct the previous solution
6. Iterate ⇒ remove interpolation
artifacts
4. Transfer error to the fine level
3. Solve for the error on the coarse level
⇒ good error estimation
Coarse grid - advantages
Coarsening allows:
make iteration process faster (on the coarse grid we can
effectively minimize the error)
obtain better initial guess u initial (solve directly on the coarsest
grid)
go to the
coarsest grid
interpolate u initial
to the finer
grid
solve here the equation
to find
u initial
Au f
Multigrid approach – Full scheme
Confidence Metric
• Intrinsic in Local Methods
• How to evaluate for global methods?
– Edge strength?
• Doesn’t work (Barron et al.,1994)
Confidence Metric
• Histogram of error contribution
Number of
pixels
Error
Confidence Metric
More Results
More Results
Further Reading
• Combining the advantages of local and global optic flow methods
(“Lucas/Kanade meets Horn/Schunck”) A. Bruhn, J. Weickert, C.
Schnörr, 2002 - 2005
• High accuracy optical flow estimation based on a theory for warping
T. Brox, A. Bruhn, N. Papenberg, J. Weickert, 2004 - 2005
• Real-Time Optic Flow Computation with Variational Methods
A. Bruhn, J. Weickert, C. Feddern, T. Kohlberger, C. Schnörr, 2003 2005
• Towards ultimate motion estimation: Combining highest accuracy
with real-time performance. A. Bruhn, J. Weickert, 2005
• Bilateral filtering-based optical flow estimation with occlusion
detection. J.Xiao, H.Cheng, H.Sawhney, C.Rao, M.Isnardi, 2006