Study with the several resources on Docsity
Earn points by helping other students or get them with a premium plan
Prepare for your exams
Study with the several resources on Docsity
Earn points to download
Earn points by helping other students or get them with a premium plan
The 'grabcut' method for efficient and interactive foreground extraction from complex backgrounds, using iterated graph cuts and a gaussian mixture model. The goal is to produce accurate and visually convincing segmentations with minimal user interaction. Performance is measured by accurate segmentation, subjective convincingness, and freedom from color bleeding. The document extends the graph cut algorithm by replacing histograms with a gaussian mixture model, using an iterative procedure, and allowing for incomplete labeling.
Typology: Study notes
1 / 26
This page cannot be seen from the preview
Don't miss anything!
By Carsten Rother, Vladimir Kolmogorovand Andrew Blake at Microsoft Research
Cambridge, UKSIGGRAPH 2004
Slides by David Anthony TorresComputer Science and Engineering — University of California at San Diego
Deal with efficient, interactive extraction of aforeground object from a complex background. ï^
Goal is to produce a ìgoodî automatic extractionwith as little user interaction as possible. ï^
Performance is measured by^!
Accurate segmentation of the object. !
Subjectively convincing extraction when faced with blur,transparency. !
Free of color bleeding in from the background.
User provides a trimap
F^
B^
U
which
partitions the image into 3 regions: foreground,background, unknown.
T^ F
T^ U
T^ B
N
N
n^
z
E(
α
, θ
, z
)^
= U
( α
, θ
, z
)^
+ V(
α
, z
)
ï^
U evaluates the fit of the opacity
α
to the data
z
!
i.e. it gives a good score (low score) if
α
looks like itís
consistent with the histogram.
ï^
V is a smoothness term which penalizes if there is toomuch disparity between neighboring pixel values.
(^
,^
,^
)^
log
(^
;^
)
n^
n
n
U
z^
h z
α θ
α
=
−
ïEdge weights are labeled with U( ) and V ( )
B
U
Each pixel
z
n^
is now in RGB color space
ï^
Color space histograms are impractical so we use aGaussian Mixture Model (GMM)^!
2 Full-covariance Gaussian mixtures with K components(K ~ 5). !
One for foreground, one for background.
ï^
Add to our model a vector
k
={ k
1
Ö k
with
k
in i^
ï^
k^ i
assigns the pixel
z
to a unique GMM component i^
(Either from F.G. or B.G. as
α
dictates)
Must incorporate
k
into our model:
E
( α
, k
, θ
, z
)^
= U
( α
, k,
θ ,
z )
+ V(
α ,
z )
where
U
( α
, k,
θ ,
z ) =
∑
n
(α
,knn
,θ
,z
)n
ï^
(α
,knn
,θ
,zn
)= ñ logn
p
(z
|n^
α
,knn
,θ
) ñ log
π
(α
,knn
)
ï^
Where
π
(∑) is a set of mixture weights which satisfy the
constraint:
k
n^
by iterating through all values 1ÖK. (K is small)
Now that
k
and
θ
are known, we can solve for the
opacity values using a minimum-cut algorithm, andreapply the min-cut until convergence. ï^
Each iteration eats away at the unknown region andcontinues to minimize the energy function.
iteration
If the segmentation isimprecise, the user canconstrain pixels to theforeground/background ï^
Then we run min-cut (orthe entire procedure)again to produce a finalsegmentation.
