Scribd
Upload
54 views

Motion Segmentation Using EM - A Short Tutorial: 1 The Expectation (E) Step

This document provides a tutorial on using the Expectation-Maximization (EM) algorithm for motion segmentation. It explains that EM can estimate both the parameters of multiple motion models (e.g. lines or translations) as well as assign data points to the appropriate model. The algorithm iterates between an E-step, where data is soft-assigned to models based on residuals, and an M-step, where the model parameters are re-estimated based on the current assignments. As an example, it describes applying EM to segment pixels into two translational motion models based on optical flow data.

Uploaded by

truongvinhlan19895148
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
54 views

Motion Segmentation Using EM - A Short Tutorial: 1 The Expectation (E) Step

This document provides a tutorial on using the Expectation-Maximization (EM) algorithm for motion segmentation. It explains that EM can estimate both the parameters of multiple motion models (e.g. lines or translations) as well as assign data points to the appropriate model. The algorithm iterates between an E-step, where data is soft-assigned to models based on residuals, and an M-step, where the model parameters are re-estimated based on the current assignments. As an example, it describes applying EM to segment pixels into two translational motion models based on optical flow data.

Uploaded by

truongvinhlan19895148
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 5

Motion Segmentation using EM - a short tutorial

Yair Weiss
MIT E10-120, Cambridge, MA 02139, USA
[email protected]

The expectation-maximization algorithm (EM) is used for many estimation problems


in statistics. Here we give a short tutorial on how to program a segmentation algorithm
using EM. Those interested in the theory or in more advanced versions of the algorithm
should consult the references at the end.
Suppose we are given a set of datapoints that were generated by multiple processes,
for example two lines. We need to estimate two things: (1) the parameters (slope and
intercept) of the two lines and (2) the assignment of each datapoint to the process that
generated it. The intuition behind EM is that each of these steps is easy assuming the
other one is solved. That is, assuming we know the assignment of each datapoint, then we
can estimate the parameters of each line by taking into consideration only those points
assigned to it. Likewise, if we know the parameters of the lines we can assign each point
to the line that ts it best.
This gives the basic structure of an EM algorithm:
 start with random parameter values for the two models.
 Iterate until parameter values converge:
{ E step: assign points to the model that ts it best.
{ M step: update the parameters of the models using only points assigned to it.
In fact both steps are slightly more complicated, due to the assignment being continuous rather than binary valued. The following sections go into more detail regarding the
E step and the M step.

1 The Expectation (E) step


In the E step we compute for each datapoint two weights w (i); w (i) (the soft assignment
of the point to models 1 and 2 respectively ). Again, we assume that the parameters
of the processes are known. Thus for each datapoint we can calculate two residuals
r (i); r (i) - the di erence between the observation at point i and the predictions of each
model. In the case of line tting the residual is simply given by:
r (i) = a xi + b yi
(1)
1

1 Throughout this tutorial we assume the number of models is known and is equal to two. A method
for automatically estimating the number of models is presented in (Weiss and Adelson, 1996)

and similarly for r (i). For example, suppose we have two line models: (1) y = x + 3
and (2) y = 2x 1. Suppose the ith datapoint is x = 1; y = 1:1. Then the residual for
line 1 is r (i) = (1 + 3 1:1) = 2:9 and for line 2 we get r (i) = (2 1 1:1 ) = 0:1 .
Intuitively we would expect this datapoint to be assigned to model 2 because the residual
is smaller. Indeed the formula for the weights is consistent with the intuition:
2

2
1

w1 (i)

2
2

r12 (i)=2

r22 (i)=2

r12 (i)=2

+e

(2)

r22 (i)=2

(3)
+ e r22 i =2
There are a few things to note about this formula. First, note that w (i) and w (i) sum
to one. This is because these weights are actually probabilities - the formula is simply
derived from Bayes rule. Second, note that there is a "free parameter" here . Roughly
speaking, this parameter corresponds to the amount of residual expected in your data
(e.g. the noise level). Finally, note that if r (i) is much smaller than r (i) then w (i) = 1
and w (i) = 0. For this reason equations 2-3 are sometimes known as the "softmin"
equations. They are a way of generalizing the concept of a minimum of two numbers
into a smooth function. More details on this can be found in the references.
So to summarize, the E step calculates two weights for every datapoint by rst calculating residuals (using the parameters of each model) and then running the residuals
through a softmin function as de ned above.
w2 (i)

r12 (i)=2

( )

2
1

2
2

2 The Maximization (M) step


In the M step we assume the weights are given, i.e. for each datapoint we know w (i) and
w (i). To estimate the parameters of each process we just use weighted least squares.
An example of least squares estimation is when we want to t a line to datapoints. It is
easy to show that the parameters of the line a; b are a solution to the following equation:
1

P x2 P x ! " a # " P x y #
Pi xi Pi 1i
= Pi yi i
b
i i

i i

Now in weighted least squares we are also given wi for each point and the equation simply
become:
"P
#
P w x2 P w x ! " a #
w
x
y
i
i
i
i
i
i
i
i
i
i
P
P
P
=
b
i wi xi
i wi 1
i wi yi

(4)

So in the M step we solve the above equation twice. First with wi = w (i) for the
parameters of line 1 and then with wi = w (i) for the parameters of line 2. In general,
in the M step we solve two weighted least squares - one for each model, with the weights
given by the results of the E step.
1

3 So what about motion segmentation?


To apply the EM algorithm for motion segmentation we need to make the following
decisions:
2

2.5

2.5

2.5

1.5

1.5

1.5

0.5

0.5

0.5

0.5

0.5

0.5

1.5

1.5

1.5

2
0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

2
0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

2
0

0.8

0.8

0.8

0.6

0.6

0.6

0.4

0.4

0.4

0.2

0.2

0
0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

0
0

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

0
0

0.8

0.8

0.6

0.6

0.6

0.4

0.4

0.4

0.2

0.2
0.1

0.2

0.3

0.4

0.5

t=1

0.6

0.7

0.8

0.9

0
0

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

0.2
0.1

0.8

0
0

0.1

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

0.2
0.1

0.2

0.3

0.4

0.5

t=2

0.6

0.7

0.8

0.9

0
0

t=3

Figure 1: An example of the EM algorithm for tting two lines to data. The algorithm
starts at random conditions and converges in three iterations. The top panels show the
line ts at every iteration and the bottom panels show the weights. (data was generated
by setting y = x +1 for jx 0:5j < 0:25 and y = x otherwise. The parameter  = 0:1).
2

 What is the class of possible motion models we are considering? What are the

parameters that need to be estimated for each model? (e.g. translational models
or ane models)
 What is the nature of the data that the models need to t? (e.g. raw pixel data or
results of optical ow).
Here we will describe the simplest version of an EM based motion segmentation algorithm:
the models will be assumed to be translational and the image data will be results of optical
ow. More complex versions are described in the references.
We are given the results of optical ow vx(i; j ) and vy (i; j ). We assume this ow was
generated by two global translation models. Each translation model is characterized by
two numbers: (u; v) which give the horizontal and vertical translation. The problem is
to estimate the parameters (u ; v ); (u ; v ) as well as assigning each pixel to the model
that generated it.
The E step is entirely analogous to the one described for line tting. We estimate two
weights for every pixel w (i; j ) and w (i; j ). This is done by assuming (u ; v ); (u ; v )
are known and calculating residuals:
r (i; j ) = (u vx (i; j )) + (v vy (i; j ))
(5)
r (i; j ) = (u vx (i; j )) + (v vy (i; j ))
(6)
The residuals are converted into weights by passing them through the softmin function:
1

2
1
2
2

w1 (i; j )
w2 (i; j )

r22 (i;j )=2

r12 (i;j )=2

+e

(7)

r22 (i;j )=2

(8)
+ e r22 i;j =2
The M step is also analogous to the one used for line tting. In this case, the
parameters (u ; v ) satisfy:
!" #
"P
#
P w (i; j )
0
u
w
(
i;
j
)
v
(
i;
j
)
x
i;j
i;j
P w (i; j )
= P w (i; j )v (i; j )
(9)
0
v
y
i;j
i;j
And the equations for (u ; v ) are the same with w (i; j ) replaced everywhere by w (i; j ).
To summarize, the motion segmentation starts by choosing an initial random value
for the translation parameters, and iterating the E and M steps detailed above until the
translation parameters converge.
In the simple case described here, since the EM algorithm is using the results of an
optical ow algorithm, it makes sense to get the nal segmentation by going back to the
raw pixel data. That is, after the algorithm has converged and two translation models
have been found, we let the models \compete" to explain the pixel data. This is done
by:
 warping frame 2 to frame 1 with the translation of model 1. Subtracting the warped
image from the true frame 1. This gives the prediction error of model 1.
 repeating the previous step with the translation of model 2 and obtaining the
prediction error of model 2.
 assigning each pixel to the model whose prediction error is lowest at that pixel.
1

r12 (i;j )=2

r12 (i;j )=2

4 Conclusion
This tutorial has discussed how to use the EM algorithm for motion segmentation. Although we have given an informal exposition, it should be noted that the actual algorithm
is derived from a rigorous statistical estimation framework and proofs about convergence
and rate of convergence can be found in the literature. The combination of the algorithm's principled derivation and intuitively decoupled steps is probably responsible for
its continued success in many estimation problems.

References
Ayer, S. and Sawhney, H. S. (1995). Layered representation of motion video using robust
maximum likelihood estimation of mixture models and mdl encoding. In Proc. Int'l
Conf. Comput. Vision, pages 777{784.
Dempster, A. P., Laird, N. M., and Rubin, D. B. (1977). Maximum likelihood from
incomplete data via the EM algorithm. J. R. Statist. Soc. B, 39:1{38.
Jepson, A. and Black, M. J. (1993). Mixture models for optical ow computation. In
Proc. IEEE Conf. Comput. Vision Pattern Recog., pages 760{761, New York.
Weiss, Y. and Adelson, E. H. (1994). Perceptually organized EM: a framework for motion
segmentation that combines information about form and motion. Technical Report
315, MIT Media Lab, Perceptual Computing Section.
Weiss, Y. and Adelson, E. H. (1996). A uni ed mixture framework for motion segmentation: incorporating spatial coherence and estimating the number of models. In Proc.
IEEE Conf. Comput. Vision Pattern Recog., pages 321{326.

You might also like