- 159 Views
- Uploaded on

Download Presentation
## PowerPoint Slideshow about 'Matting' - fairfax

**An Image/Link below is provided (as is) to download presentation**

Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author.While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server.

- - - - - - - - - - - - - - - - - - - - - - - - - - E N D - - - - - - - - - - - - - - - - - - - - - - - - - -

Presentation Transcript

OutlineOutlineOutlineOutline

Outline

- The matting problem
- Previous work
- New approaches:
- The iterative approach Jue Wang,Michael F.Cohen
- Closed form solutionAnat Levin, Dani Lischinski,Yair Weiss
- Comparison and summary
- Bonus?

Outline

- The matting problem
- Previous work
- New approaches:
- The iterative approach Jue Wang,Michael F.Cohen
- Closed form solutionAnat Levin, Dani Lischinski,Yair Weiss
- Comparison and summary
- Bonus?

The matting problem

- The separation of an image I into
- Foreground object image F
- Background image B
- Alpha matte α – the opacity

- Problem: extract F, B, α from image

hair

fur

Why is matting challenging?

- Under constrained problem:One equation, 3 unknowns

We need to constrain the problem!

Outline

- The matting problem
- Previous work
- New approaches:
- The iterative approach Jue Wang,Michael F.Cohen
- Closed form solutionAnat Levin, Dani Lischinski,Yair Weiss
- Comparison and summary
- Bonus?

Known Background

- Blue screen Matting
- Still under-constrained
- Solution: make more assumptions
- “Foreground contains no blue”
- Other foreground distribution assumption…
- Use two different backgrounds
- Main flaw: need to know the background…

Blue background

Composite image

Natural Image Matting

- The assumptions:
- Smoothness of the alpha matte
- GMM for the Background and Foreground colors
- Initial estimate:trimap provided by the user

Background

Foreground

Unknown

Input image

Trimap

Natural Image Matting

- The algorithms framework:
- Estimate F, B distributions from close pixels
- Find best α by some method

Knockout

- Extrapolate F,B from close neighborhood
- Estimate α from calculated F, B values

Bayesian

- Estimate F, B distributions in area
- Find best α matching distributions

Bayesian

- P(F), P(B) from image samples
- P(C|F,B,α) using a distribution for C

Natural Image Matting

- Main flaw: Accurate trimap required
- Tedious to provide manually

- Hard to extract automatically

In particular, not feasible to videos

Input image

Trimap

Binary segmentation

Adding unknown region

- The matting problem
- Previous work
- New approaches:
- The iterative approach Jue Wang,Michael F.Cohen
- Closed form solutionAnat Levin, Dani Lischinski,Yair Weiss
- Comparison and summary
- Bonus?

New Approach to Matting

Trimap reduces to scribbles

Two new methods

- Iterative optimization approach
- Heuristic algorithmic optimization
- A closed form solution
- Mathematical approach

Trimap

Scribbles

Iterative approach

- Score:

fit to image data +alpha matte smoothness

- Iteratively propagating estimated results.

Iterative optimization - outline

- Initialize “work pixels” from scribbles
- Repeatedly:
- Expand work pixels
- Find best alpha matte
- Stop when finished

ui = 0

α = 0

ui = 0

α = 1

ui = 1

α = 0.5

Initialization- Introducing:
- ui - uncertainty variable
- Uc – work pixels

Optimization

Uc = {user scribbles + 15 pixel radius}

Our goal:

find α matte for Uc that minimizes the energy -

Smoothness

Data

Vd

- Fit measure of αp to Ip
- Score for αp = α :

Fi , Bj – possible values for F, B in the pixel

wFi, wBj – corresponding weights

Vs

- Matte smoothness :

Iterative optimization – step 2

Uc = {user scribbles + 15 pixel radius}

Our goal: find α matte for Uc that minimizes the energy -

Uc Graph

Nodes = Pixels, Edges by 4-connectivity

Iterative optimization – step 2

GOAL: Minimize

BELIEF PROPAGATION

t=0

y

mpq – message from p to q

q

p

Vector: p’s “opinion” for each

possible α for q

Iterative optimization – step 2

GOAL: Minimize

BELIEF PROPAGATION

t=1

y

mpq – new message pq

myp – previous message yp

q

p

Iterative optimization – step 2

GOAL: Minimize

BELIEF PROPAGATION

t=T (stopping time)

y

q

p

Best state calculated for each node:

Iterative optimization – step 3

Found α matte for Uc that minimizes the energy -

Update F, B and uncertainty:

Iterative optimization - algorithm

- Initialize Uc, F, B, u and alpha matte from scribbles
- Repeatedly:
- Expand Uc by another 15 pixel radius
- Find best alpha matte (BP)
- Update F,B,u for new matte
- Stop when total uncertainty is minimal

Initial matte

Propagation of α matte

Final matte

Iterative optimization - Results

The ambiguity bunny

Iterative optimization - Results

Scribbles result

Trimap result

Ambiguity bunny with trimap

Ambiguity bunny with scribbles

Iterative optimization - Summary

- Minimal user input
- Applicable to video
- Sensitive to ambiguity in F, B
- Uses simple color-model
- Performance:
- 15-20 min. on a 640x480 image
- Factor 50 reported by better implementation

- The matting problem
- Previous work
- New approaches:
- The iterative approach Jue Wang,Michael F.Cohen
- Closed form solutionAnat Levin, Dani Lischinski,Yair Weiss
- Comparison and summary
- Bonus?

Closed form solution

- Assumption: local smoothness in F, B

cancel out unknowns from the matte equs.

- Solve for F,B and alphausing algebraic tricks.

Closed form solution

- GOAL:
- Minimize:

Closed form solution

- Minimize:

3N Variables (N = image size) We can rid a, b by algebraic manipulation

Closed form solution

- Minimize:

Proof: Rewrite in matrix form:

By mean-least-squares, best a,b pair

for each window is:

Closed form solution

- Some more manipulation give the required result

EXCITED?

GET YOUR I LOVE MATH

T-SHIRT, NOW FOR ONLY $1999

F2

F1

G

Closed form solution- For color images:
- Simple: Do each channel separately
- Smart: Assume one alpha for R,G,B.Use redundancy to allow a “color-line” model per window

Color line model:

OUT: F, B Constant within a window

IN: F, B are on some line

Closed form solution

- For color images:
- Simple: Do each channel separately
- Smart: Assume one alpha for R,G,B.Use redundancy to allow a “color-line” model per window

Closed form solution

- For color images:
- Simple: Do each channel separately
- Smart: Assume one alpha for R,G,B.Use redundancy to allow a “color-line” model per window

Now, as before, cost is:

And a,b can be cancelled out.

Closed form solution

Now problem reduced to finding best α for:

L is Huge size NxN (N = # image pixels)

But Sparse…

Closed form solution

- The algorithm:
- Compute L
- Solve for given the scribbles.
- Solving a sparse set of bilinear equationswith constraints (Lagrange multipliers)
- Find F, B given the matte
- Adding smoothness assumptions on F, B
- Improvements:
- Use larger environment in low cost by “pyramids”

Eigenvectors as guides

Small eigenvectors of L are

correlated with minimal matte

L is positive definite.

Eigenbasis: v1,…,vN

Eigenvalues: λ1 > λ2 > … > λN > 0

Eigenvectors as guides

Small eigenvectors of L are

correlated with minimal matte

can guide user scribbles

Eigenvectors matching smallest eigenvalues

Resulting matte

Guided scribbles

Closed form solution - Summary

- Minimal user input
- Provable optimality (under assumptions)
- Assumes only smooth F,B (no color model)
- Applicable to video (as we speak…)
- Problematic with textures
- Performance:
- 20-40 seconds for a 200x300 image
- Expensive in memory

- The matting problem
- Previous work
- New approaches:
- The iterative approach Jue Wang,Michael F.Cohen
- Closed form solutionAnat Levin, Dani Lischinski,Yair Weiss
- Comparison and summary
- Bonus?

Comparison

Color ambiguity

Iterative approach Closed form

Sensitive Sensitive

Solvable by adding more scribbles

Comparison

Optimality?

Iterative approach Closed form

Provably optimal

But for the specific

(simplified) cost

Uses heuristics

to optimize

Comparison

Textures

Iterative approach Closed form

F,B must satisfy

color-line model

Assumes only

Alpha matte smooth

Comparison

Rough edges

Iterative approach Closed form

Input image with scribbles

Can handle rough

edges

Assumes

Alpha matte smooth

matte results

Comparison

Running time

Iterative approach Closed form

20/40 seconds

Costly in memory

~20 sec.

(For medium size image)

Comparison

Tests

Iterative approach Closed form

Extensively tested

quantitative results

No quantitative

results reported

- The matting problem
- Previous work
- New approaches:
- The iterative approach Jue Wang,Michael F.Cohen
- Closed form solutionAnat Levin, Dani Lischinski,Yair Weiss
- Comparison and summary
- Bonus?

Environment Matting and Compositing

Douglas E. Zongker ~ Dawn M. Werner ~ Brian Curless ~ David H. Salsin

Environment Matting

C = F + (1- a)B + F

- F ~ Contribution of light from Environment that travels through the object

R – reflectance image

T – Texture image

Environment Matting?

Alpha Matte Environment Matte Photograph

Environment Mattin

Alpha Matte Environment Matte Photograph

Summary

- The matting problem
- Old methods: require trimap
- Two new methods from scribbles:
- Iterative optimization
- Assume: matte smooth, F,B locally similar
- Use heuristic optimization for alpha
- Close form solution
- Assume: F, B locally smooth (color-line model)
- Solve linear equations for alpha

Download Presentation

Connecting to Server..