Structure of Trackingbydetection with Kernels Seunghoon Hong CV Lab POSTECH Motivation Trackingbydetection A classifier is trained in online using examples patches obtained during tracking ID: 716961
Download Presentation The PPT/PDF document "Exploiting the Circulant" is the property of its rightful owner. Permission is granted to download and print the materials on this web site for personal, non-commercial use only, and to display it on your personal computer provided you do not modify the materials and that you retain all copyright notices contained in the materials. By downloading content from our website, you accept the terms of this agreement.
Slide1
Exploiting the Circulant Structure of Tracking-by-detection with Kernels
Seunghoon Hong
CV Lab.
POSTECHSlide2
MotivationTracking-by-detection
A classifier is trained
in on-line using examples (patches) obtained during trackingSlide3
MotivationSampling strategy
Sparse sampling for computational efficiencySlide4
MotivationRedundancy in samplingA set of dense patches has extreme redundanciesSlide5
MotivationKey ideaRepresenting a set of dense samples by
“
circulant
structure”
Training and testing can be extremely efficient by exploiting such structureSlide6
Circulant matrix
Circulant matrix
Concatenation of all possible cyclic shifts of
Slide7
Circulant matrix
Benefit of
circulant
matrix
Only a base vector
need to be stored
Inner product can be computed very efficiently
Convolution between
and
Slide8
Circulant matrix
Benefit of
circulant
matrix
Only a base vector
need to be stored
Inner product can be computed very efficiently
Convolution between
and
Efficiently computed by
Fast Fourier transform
!Slide9
Circulant matrix
Benefit of
circulant
matrix
Only a base vector
need to be stored
Inner product can be computed very efficiently
It
can be generated by permutation matrix
row of
is obtained by
Slide10
Circulant matrix
Representing dense samples by
circulant
matrix
Given a base sample
,
generate a (virtual) dense sample matrix
by
shifting
direction
Slide11
Learning with dense sampling
Regularized risk minimization
Given a set of sample
,
optimize a classifier
by
where
is loss function and
is regularization parameter
Slide12
Learning with dense sampling
Regularized risk minimization
Kernel trick
: map inputs to feature space
by
Representer
theorem
: the solution can be obtained by linear combination of inputs by
Close form solution of KRLS
Kernel matrixSlide13
Learning with dense sampling
Kernel matrix with dense sampling
Kernel matrix
is
circulant
if
is unitarily invariant kernel
(e.g. linear, RBF, polynomial)
is unitarily invariant if
for any unitarily matrix
=
Slide14
Learning with dense sampling
Efficient kernel regularized RLS solution
Define kernel vector
with elements
where
Solution of KRLS is rewritten as
Slide15
Learning with dense sampling
Fast detection
Given a test image
, a testing is performed by
Then it can be re-written by
, where
Applying property of
circulant
matrix,
Slide16
Learning with dense sampling
Applying different kernels
Let
then
Slide17
Learning with dense sampling
Applying different kernels
Linear kernel
Slide18
Learning with dense sampling
Applying different kernels
Linear kernel
Polynomial kernel
Slide19
Learning with dense sampling
Applying different kernels
Linear kernel
Polynomial kernel
RBF kernel
Slide20
Overall algorithmSlide21
Experiments
Preprocessing
To reduce boundary effects.
Band original patch
with a cosine window.
Original image
After preprocessingSlide22
ExperimentsTraining label
Smooth regression output by Gaussian kernel
Smoothed output
Slide23
ExperimentsSlide24
Q & A