Optical flow

Or where do the pixels move?

Alon Gat

Given: two or more frames of an image sequence

Wanted: Displacement field between two consecutive frames

optical flow

Problem Definition

Visualize Optical Flow

Vector Plot: Subsample vector field and use arrows for visualization

Color Plot: Visualize direction as color and magnitude as brightness

Extraction of Motion Information • robot navigation/driver assistance • surveillance/tracking • action recognition

Processing of Image Sequences • video compression • ego motion compensation

Related Correspondence Problems • stereo reconstruction • structure-from-motion • medical image registration

What is optical flow good for?

• How to estimate pixel motion from two images?– Find pixel correspondences

• Given a pixel in img1, look for nearby pixels of the same color in img2

• Key assumptions– color constancy: a point in img1 looks “the same” in

img2• For grayscale images, this is brightness constancy

– small motion: points do not move very far

Assume brightness of patch remains same in both images:

),( yx

),( tvytux

ttime tttime ),( yx

Optical Flow: the vector field

),( vu

Displacement: ),(),( tvtuyx

),,(),,( tyxItttvytuxI

Brightness Constancy Assumption

The Linearized Brightness Constancy Assumption

),,(),,( tyxItttvytuxI

),,(),,( tyxIt

Ivy

Iux

ItyxI

Brightness Constancy Assumption

Idea: If u and v are small and I is sufficiently smooth, one may linearize thisconstancy assumption via a first-order Taylor expansion around the point

0 tyx IvIuI Known

Unknown

Image derivative

Smoothness Constraint :meaning neighbor pixels in the picture has similar velocities.

Horn & Schunck Algorithm (MIT 1981)

,))()(( 2222 dxdyvvuu yxyx

,22dxdyvues

In other words, nearby pixels moves together

We seek the set that minimize:

Horn & Schunck Algorithm

dxdyvvuuIvIuIyxvyxuE yxyxtyx ))()(()),(),,(( 22222

),(),,( yxvyxu

Smoothness term

Data term brightness constancy

• data term - penalizes deviations from constancy assumptions• smoothness term - penalizes dev. from smoothness of the solution• regularization parameter α - determines the degree of smoothness

Output – the optical flow!

SmoothingIdea: In order to reduce the influence of noise and outliers, we convolve I0 with a Gaussian of mean μ = 0 and standard deviation

),,(),,( 0 tyxItyxI

Gaussian

Horn & Schunck Algorithm

According to the calculus of variations, a minimizer of E must fulfill the Euler-Lagrange equations

Which are highly non linear system of equations…

0

yx uuu F

yFx

F

dxdyvvuuvuyxFvuE yxyx ),,,,,,,(),(

Euler-Lagrange equations

0

yx vvv F

yFx

F

Calculus of Variation

Euler-Lagrange equations

dxdyvvuuIvIuIyxvyxuE yxyxtyx ))()(()),(),,(( 22222

)()(0

)()(0

yyxxytyx

yyxxxtyx

vvIIvIuI

uuIIvIuI

2

2

2

2

yx

ytyx

xtyx

IIvIuIv

IIvIuIu

)(

)(

Or

So we linearize again!

Horn & Schunck Algorithm

ytyx

xtyx

IIvIuIv

IIvIuIu

)(

)(

2

,1,

2

,1,

2

,,1

2

,,1

x

jiji

x

jiji

x

jiji

x

jiji

h

uu

h

uu

h

uu

h

uuu

flow derivatives here discredited via

linear system of equations

Horn & Schunck Algorithm

0)()( klklklx

kltkl

klykl

klx uuIIvIuI

0)()( klklkly

kltkl

klykl

klx vvIIvIuI

klt

klykl

klt

klxkl

kl

klkly

klx

kly

kly

kly

klx

klx

klx

IIv

IIu

v

u

IIII

IIII

Horn & Schunck Algorithm

klt

klykl

klt

klxkl

kl

klkly

klx

kly

kly

kly

klx

klx

klx

IIv

IIu

v

u

IIII

IIII

klxkl

yklx

klt

nkl

kly

nkl

klxn

klnkl I

II

IvIuIuu

])()[( 221

klykl

yklx

klt

nkl

kly

nkl

klxn

klnkl I

II

IvIuIvv

])()[( 221

Update Rule:

Examples

Output

Hard to find boundaries.

Two approximations. Less accurate.

Problems with the method.

Instead of approximating the brightness constancy to the 1st Tylor expansion, we’ve add one order.

Second and more important, instead of solving E-L equations (which needed to be linearized) we wrote the Functional as n*m equations and minimized it with regular minimization methods (Gradient decent, Quasi Newton, and others)

So what are we trying to do?

1 .Mathemactica for symbolic calculations.

Where P are the Image derivatives, and u, v is the optical flow

Mathematica Symbolic Toolbox for MATLAB--Version 2.0(http://library.wolfram.com/infocenter/MathSource/5344/)

Quasi Newton IterationThe problem was that calculation time of inverse of non sparse matrix was long.And then multiplying two non sparse matrix…

Gradient Decent.Non of the above problems but linear convergence rate.And convergence to local min.

2 .Matlab for numeric calculations

In order to get things going faster (moving loooooooong string from matlab to mathematica takes awhile), we found the functional matrix’s constancy and calculate it in matlab.

for i=2:imageSizeN-1 for j=2:imageSizeM-1 gradC(i,j)=gradC(i,j)+alpha*(2*(-c(i-1,j)+c(i,j))+2*(-c(i,j-1)+c(i,j))-2*alpha*(-c(i,j)+c(i,j+1)) - 2*alpha*(-c(i,j)

+c(i+1,j))+4*exp((-1+c(i,j)^2+s(i,j)^2)^2)*c(i,j)*(-1+c(i,j)^2+s(i,j)^2) +2*(Ix(i,j)*m(i,j)+Ixz(i,j)*m(i,j)+2*Ixx(i,j)*c(i,j)*m(i,j)^2+Ixy(i,j)*m(i,j)^2*s(i,j))*(Iz(i,j)+Izz(i,j)+Ix(i,j)*c(i,j)*m(i,j)+Ixz(i,j)*c(i,j)*m(i,j)+Ixx(i,j)*c(i,j)^2*m(i,j)^2+Iy(i,j)*m(i,j)*s(i,j) +Iyz(i,j)*m(i,j)*s(i,j)+Ixy(i,j)*c(i,j)*m(i,j)^2*s(i,j)+Iyy(i,j)*m(i,j)^2*s(i,j)^2));

gradS(i,j)=gradS(i,j)+4*exp((-1+c(i,j)^2+s(i,j)^2)^2)*s(i,j)*(-1+c(i,j)^2+s(i,j)^2)+2*(Iy(i,j)*m(i,j)+Iyz(i,j)*m(i,j)+Ixy(i,j)*c(i,j)*m(i,j)^2+2*Iyy(i,j)*m(i,j)^2*s(i,j))*(Iz(i,j)+Izz(i,j) +Ix(i,j)*c(i,j)*m(i,j)+Ixz(i,j)*c(i,j)*m(i,j)+Ixx(i,j)*c(i,j)^2*m(i,j)^2+Iy(i,j)*m(i,j)*s(i,j)+Iyz(i,j)*m(i,j)*s(i,j)+Ixy(i,j)*c(i,j)*m(i,j)^2*s(i,j)+Iyy(i,j)*m(i,j)^2*s(i,j)^2) +alpha*(2*(-s(i-1,j)+s(i,j))+2*(-s(i,j-1)+s(i,j))-2*alpha*(-s(i,j)+s(i,j+1))-2*alpha*(-s(i,j)+s(i+1,j)));

gradM(i,j)=gradM(i,j)+beta*(2*(-m(i-1,j)+m(i,j))+2*(-m(i,j-1)+m(i,j)))-2*beta*(-m(i,j)+m(i,j+1))-2*beta*(-m(i,j)+m(i+1,j))+2*(Ix(i,j)*c(i,j)+Ixz(i,j)*c(i,j) +2*(Ixx(i,j)*c(i,j)^2*m(i,j)+Iy(i,j)*s(i,j)+Iyz(i,j)*s(i,j)+2*Ixy(i,j)*c(i,j)*m(i,j)*s(i,j)+2*Iyy(i,j)*m(i,j)*s(i,j)^2)*(Iz(i,j)+Izz(i,j)+Ix(i,j)*c(i,j)*m(i,j)+Ixz(i,j)*c(i,j)*m(i,j) +Ixx(i,j)*c(i,j)^2*m(i,j)^2+Iy(i,j)*m(i,j)*s(i,j)+Iyz(i,j)*m(i,j)*s(i,j)+Ixy(i,j)*c(i,j)*m(i,j)^2*s(i,j)+Iyy(i,j)*m(i,j)^2*s(i,j)^2));

end end end

3 .Leaving Mathematica

Gradient Decent

All methods need couple of unknown parameters which need to be selected by an educated guess. (Condor to the rescue)

All the image derivatives and gradients are calculated in a linear manner.

So what is there to optimize?

First results

Takes about 45min(highend algorithms take around 20sec…)

The Ground Truth.

Questions?