1. GrabCut Interactive Image (and Stereo) Segmentation Joon Jae Lee Keimyung University
Welcome. I will present Grabcut – an Interactive tool for foreground extraction.

2. Characteristics Improved from graph cut
– Use Gaussian Mixture Model (GMM) for clustering in color space
– Iterative optimization
User interaction greatly reduced compared
with other methods (Magic Wand, Intelligent
Scissors)

3. Photomontage GrabCut – Interactive Foreground Extraction 1
Here are some images from my last hiking trip in England. And would it be great if we could just do the following. Drag rectangle. Create a nice photomontage with an extreamly simple user interface.
GrabCut – Interactive Foreground Extraction

4. Problem Fast & Accurate ?
State the problem. Image object to extract and simple tool which gives high quality results with an alpha matte and even works for the case of camoflage.
GrabCut – Interactive Foreground Extraction

5. Intelligent Scissors Mortensen and Barrett (1995)
What GrabCut does
Magic Wand (198?)
Intelligent Scissors Mortensen and Barrett (1995)
GrabCut
User Input
Result
Basic idea is to combine the information which was used in the 2 well known tools: MW and IS.
MS select a region of similar colour according to your input. Fat pen and the boundary snapps to high contrast edges.
Grabcut combines it. And this allows us to simplify the user interface considerably by …
Regions
Boundary
Regions & Boundary
GrabCut – Interactive Foreground Extraction

6. Gaussian Mixture Model
The probability of a pixel vector is represented
as:

7. Gaussian Mixture Model
Two sets of GMM, one for background, one
for unknown region

8. Algorithm For each pixel assign a value α
α = 0 for background and α = 1 for foreground
Use graph cut to minimize a Gibbs energy:
α: background / foreground label
θ: GMM parameters
k: GMM labeling
z: pixel value

9. Iterative minimization
Initialize background pixel to α = 0 and unknown region (draw box) to α = 1.
Initialize two sets of GMMs.
1. Assign GMM labels to each pixel. (Which set of GMM is determined by α)
2. Graph cut minimize (α optimized, GMM label changed to corresponding set of GMM)
3. Update GMM parameters
4. Repeat from step 1 until convergence

10. Graph Cuts - Boykov and Jolly (2001)
Foreground (source)
Image
Min Cut
Cut: separating source and sink; Energy: collection of edges
Min Cut: Global minimal enegry in polynomial time
Optimization engine we use Graph Cuts. By know everybody should know what graph cut is. IN case you missed it here is a very brief introduction.
To image. 3D view. First task to construct a graph.
All pixels on a certain scanline. Just a few of them.
Next step introduce artificial nodes. fgd and background.
Edges – contrast. Boundary is quite likely between black and white.
Min Cut - Minimum edge strength.
Background (sink)
GrabCut – Interactive Foreground Extraction

11. Guaranteed to converge
Iterated Graph Cuts
Guaranteed to converge
Iterations are not shown to the user; Converges: proof in the paper. 1 2 3 4
Result
Energy after each Iteration
GrabCut – Interactive Foreground Extraction

12. Colour Model R R G G Gaussian Mixture Model (typically 5-8 components)
Iterated graph cut
Foreground & Background
Foreground
Color enbergy.
Iterations have the effect of pulling them away;
D is – log likelyhood of the GMM.
Background G
Background G
Gaussian Mixture Model (typically 5-8 components)
GrabCut – Interactive Foreground Extraction

13. Coherence Model An object is a coherent set of pixels: 25
Also coherent model. Strength of contrast – colour difference. Lambda importance of coherence model. 25
Error (%) over training set:
How do we choose ? 25

14. Parameter Learning - Problems
A Gaussian MRF is not a realistic texture model
syntheticGMRF
Gaussian?
Real Image
Gaussian!
GrabCut – Interactive Foreground Extraction

15. Moderately simple examples
Moderately straightforward examples- after the user input automnatically
… GrabCut completes automatically
GrabCut – Interactive Foreground Extraction

16. Difficult Examples Camouflage & Low Contrast Fine structure
No telepathy
Initial Rectangle
Initial Result
You might wonder when does it fail. 4 cases. Low contrats – an edge not good visible
GrabCut – Interactive Foreground Extraction

17. Evaluation – Labelled Database
Labeled data base 70 images . availoable online. Different scenarios – simple and ifficult shapes and colour.
Available online:
GrabCut – Interactive Foreground Extraction

18. Comparison Boykov and Jolly (2001) GrabCut Error Rate: 0.72%
User Input
Error Rate: 0.72%
Databsed used to compare to BJ. Re-implented their method. Simpler same error rate
Result
Error Rate: 0.72%
Error Rate: 1.32%
Error Rate: 1.87%
Error Rate: 1.81%
Error Rate: 1.25%
GrabCut – Interactive Foreground Extraction

19. Trimap Boykov and Jolly
Comparison
Input Image Ground Truth
Trimap Boykov and Jolly
Error Rate: 1.36%
Bimap GrabCut
Error Rate: 2.13%
Use our database. Small uncertainty area.
Error rate - modestly increase
User Interactions - considerable reduced
GrabCut – Interactive Foreground Extraction

20. Results Parameter Learning
GrabCut – Interactive Foreground Extraction

21. Comparison Magic Wand (198?)
Finally – number of tools: certain scenarios.
what does the user want to do for segmentation –
Graph Cut – user input reduced and quality increases.
Magic Wand (198?)
Intelligent Scissors Mortensen and Barrett (1995)
Graph Cuts Boykov and Jolly (2001)
LazySnapping Li et al. (2004)
GrabCut Rother et al. (2004)
GrabCut – Interactive Foreground Extraction