OPPA European Social Fund Prague & EU: We invest in your ... · OPPA European Social Fund Prague &...
Transcript of OPPA European Social Fund Prague & EU: We invest in your ... · OPPA European Social Fund Prague &...
OPPA European Social FundPrague & EU: We invest in your future.
Talk Outline
� appearance based tracking
� patch similarity using histogram
� tracking by mean shift
� experiments, discussion
Mean shift 1
Tomáš Svoboda, [email protected] Technical University in Prague, Center for Machine Perception
http://cmp.felk.cvut.czLast update: April 8, 2013
1Please note that the lecture will be accompanied be several sketches and derivations on the blackboardand few live-interactive demos in Matlab
2/21Appearance based tracking
2
2illustration from [1]
3/21Histogram based representation
image patch
5 10 15 20 25
5
10
15
20
25
30
0 50 100 150 200 2500
50
100
150
200
250
300
intensity
freq
uenc
y
histogram for 16 bins
4/21Patch comparison
image patch
5 10 15 20 25
5
10
15
20
25
30
0 50 100 150 200 2500
50
100
150
200
250
300
intensity
frequ
ency
histogram for 16 bins
image patch
5 10 15 20 25
5
10
15
20
25
50 100 150 200 2500
10
20
30
40
50
60
70
80
intensity
frequ
ency
histogram for 16 bins
5/21histogram difference
assume normalized histograms, i.e∑mu=1 pu = 1
d =√1− ρ[p, q]
where ρ[p, q] is the Bhattacharyyacoefficient
ρ[p, q] =m∑
u=1
√puqu
image patch
5 10 15 20 25
5
10
15
20
25
30
0 50 100 150 200 2500
50
100
150
200
250
300
intensity
frequency
histogram for 16 bins
image patch
5 10 15 20 25
5
10
15
20
25
50 100 150 200 2500
10
20
30
40
50
60
70
80
intensity
frequency
histogram for 16 bins
6/21
Similarity measured by the Bhattacharyyacoefficient
Similarity surface of the frame 206
100 200 300 400 500 600 700
50
100
150
200
250
300
350
400
450
500
5500
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
The object is the “4” plate and the model is histogram of image intensities.
s(y) =m∑
u=1
√pu(y)qu
where p(y) is the histogram of image patch at position y and q is thehistogram of the template.
7/21Problem: finding modes in probability density
Similarity surface of the frame 206
100 200 300 400 500 600 700
50
100
150
200
250
300
350
400
450
500
5500
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
� the complete enumeration of similarity surface can be costly,
� can we do it faster and more elegantly?
8/21
Density Gradient Estimation
9/21Mean shift procedure
3
3Figure borrowed from [4]
10/21
Meanshift segmentation of colours - colordistribution
4
4Figure from [2]
11/21
Meanshift segmentation of colours - color modesseeking
5
5Figure from [2]
12/21Mean shift segmentation - intensity and space
u, v are here spatial pixel coordinates
different normalization for intensity and spatial coordinates
13/21Multivariate kernel density estimator
−1−0.5
00.5
1
−1
−0.5
0
0.5
10.2
0.4
0.6
0.8
1
Normal kernel, KN (x) = exp(−12‖x‖2)
−1−0.5
00.5
1
−1
−0.5
0
0.5
1−1
−0.5
0
0.5
1
Epanechnikov kernel, KE(x) = 1 −‖x‖2
Given n data points xi in d-dimensional space Rd.
f̃h,K(x) =1
nhd
n∑
i=1
K
(x− xi
h
)
� looking for extremum of fh,K(x)
� gradient ∇fh,K(x) = 0
14/21Kernels
−1−0.5
00.5
1
−1
−0.5
0
0.5
10.2
0.4
0.6
0.8
1
Normal kernel, KN (x) = exp(−12‖x‖2)
−1−0.5
00.5
1
−1
−0.5
0
0.5
1−1
−0.5
0
0.5
1
Epanechnikov kernel, KE(x) = 1 −‖x‖2
Can be seen as membership function.Remind Kernel density estimation (Parzen method).
Remember convolution? 6
6Taken from http://en.wikipedia.org/wiki/Kernel_density_estimation
15/21Mean-shift iterations
Assuming a reasonable differentiable kernel K, iterate till convergence:
yk+1 =
∑ni=1 xig(‖yk − xi‖2)∑n
i=1 g(‖yk − xi‖2)
g is the derivative of kernel profile.
16/21Mean-shift tracking - ratio histogram
Ratio histogram:
ru = min
(qu
pu, 1
)
where q is the histogram of the targetand p is the histogram of the currentframe. wi = rb(xi) (just binning)
Image intensities (or colors) are tranformed into weights, wi, by backprojection of the ratio histogram. Mean-shift iterations:
yk+1 =
∑ni=1wixig(‖yk − xi‖2)∑n
i=1wig(‖yk − xi‖2)
17/21Mean-shift tracking - Bhattacharya coeficient
model, coordinates x∗i centered at 0:
qu = Cn∑
i=1
k(‖x∗i‖2)δ(b(x∗
i )− u)
target candidate centered at y:
pu(y) = Ch
nh∑
i=1
k
(∥∥∥∥y − xi
h
∥∥∥∥2)δ(b(xi)− u)
18/21ms tracking - object and its model
7
7Figure from [5]
19/21ms tracking - iterations
8
8Figure from [5]
20/21References
Mean-shift originally from [3].
[1] Robert Collins. CSE/EE486 Computer Vision I. slides, web page.http://www.cse.psu.edu/~rcollins/CSE486/. Robert kindly gave general permission to reuse thematerial.
[2] Dorin Comaniciu and Peter Meer. Mean shift: A robust approach toward feature space analysis. IEEETransactions on Pattern Analysis and Machine Analysis, 24(5):603–619, May 2002.
[3] Keinosuke Fukunaga and Larry D. Hostetler. The estimation of the gradient of a density function, withappilcations in pattern recognition. IEEE Transactions on Information Theory, 21(1):32–40, January 1975.
[4] Milan Šonka, Václav Hlaváč, and Roger Boyle. Image Processing, Analysis and Machine Vision.Thomson, 3rd edition, 2007.
[5] Tomáš Svoboda, Jan Kybic, and Václav Hlaváč. Image Processing, Analysis and Machine Vision. AMATLAB Companion. Thomson, 2007. Accompanying www site http://visionbook.felk.cvut.cz.
21/21End
OPPA European Social FundPrague & EU: We invest in your future.