ron Numerical Geometry of Images: Shape Reconstruction Ron Kimmel Geometric Image Processing Lab...

16
www.cs.technion.ac.il/~ ron Numerical Geometry of Images: Shape Reconstruction Ron Kimmel Geometric Image Processing Lab Computer Science Department Technion-Israel Institute of Technolog
  • date post

    19-Dec-2015
  • Category

    Documents

  • view

    218
  • download

    0

Transcript of ron Numerical Geometry of Images: Shape Reconstruction Ron Kimmel Geometric Image Processing Lab...

Page 1: ron Numerical Geometry of Images: Shape Reconstruction Ron Kimmel Geometric Image Processing Lab Computer Science Department Technion-Israel.

www.cs.technion.ac.il/~ron

Numerical Geometry of Images:

Shape ReconstructionRon Kimmel

Geometric Image Processing Lab

Computer Science Department Technion-Israel Institute of Technology

Page 2: ron Numerical Geometry of Images: Shape Reconstruction Ron Kimmel Geometric Image Processing Lab Computer Science Department Technion-Israel.

Shape reconstruction techniques

Shape from: Inputs

Shading

Stereo

Auto-stereograms

Photometric stereo

Structured/coded light

IR-pulses (3DV-Systems)

Page 3: ron Numerical Geometry of Images: Shape Reconstruction Ron Kimmel Geometric Image Processing Lab Computer Science Department Technion-Israel.

Stereo

Page 4: ron Numerical Geometry of Images: Shape Reconstruction Ron Kimmel Geometric Image Processing Lab Computer Science Department Technion-Israel.

Shape from Stereo

Page 5: ron Numerical Geometry of Images: Shape Reconstruction Ron Kimmel Geometric Image Processing Lab Computer Science Department Technion-Israel.

Shape from autostereograms

for i = (stripesize+1):xsize+stripesize, for j = 1:ysize, stereo(j,i) = stereo(j,i-floor(stripesize*(1-z(j,i)))); end %for jend %for i

Page 6: ron Numerical Geometry of Images: Shape Reconstruction Ron Kimmel Geometric Image Processing Lab Computer Science Department Technion-Israel.

Image Formation

F. Guichard 93Mondrian world:Lambertian surface patches

lyxNyxyxI

),,(),(),(

Page 7: ron Numerical Geometry of Images: Shape Reconstruction Ron Kimmel Geometric Image Processing Lab Computer Science Department Technion-Israel.

Image formationLambetian

model

V

lN

)cos(

,),(

lNyxI

Page 8: ron Numerical Geometry of Images: Shape Reconstruction Ron Kimmel Geometric Image Processing Lab Computer Science Department Technion-Israel.

Lambertian surfaces

Images formation

.1

1),( ˆfor

,),,(),(

2z

yxIzl

lyxNyxI

Page 9: ron Numerical Geometry of Images: Shape Reconstruction Ron Kimmel Geometric Image Processing Lab Computer Science Department Technion-Israel.

Shape from photometric stereo

Given Lambertian shading model, multiple images each with different light source.

At each point we have 3 unknowns that can be computed from at least 3 images.

Next, the problem is surface integration from its gradient vector field

.1

}1,,{

,),,(),(),(

2z

zzN

lyxNyxyxI

yx

ii

,,, yx zz

},{ yx zzz

Page 10: ron Numerical Geometry of Images: Shape Reconstruction Ron Kimmel Geometric Image Processing Lab Computer Science Department Technion-Israel.

Shape from photometric stereo

Next, the problem is surface reconstruction from

Define the integral measure

For which the Euler-Lagrange is a Poisson equation

or in more compact notations

dxdyqzpzzE yx 22 )()()(

},{ qpz

yxyyxx qpzz

),( yxfz

Page 11: ron Numerical Geometry of Images: Shape Reconstruction Ron Kimmel Geometric Image Processing Lab Computer Science Department Technion-Israel.

Shape from photometric stereo

Efficient numerical algorithm for solving a Poisson equation depends on the boundary conditions.

For periodic b.c. Fourier transform is an option

),( yxfz

22

1

2222

),(ˆ),(

)()()(

)()()()),((

)),((),(ˆ

)),((),(ˆ

vu

vufyxz

zvuzvzu

zzzzyxz

yxfvuf

yxzvuz

yyxxyyxx

F

FFF

FFFF

F

F

Page 12: ron Numerical Geometry of Images: Shape Reconstruction Ron Kimmel Geometric Image Processing Lab Computer Science Department Technion-Israel.

Shape from photometric stereo

Efficient numerical algorithm for solving a Poisson equation depends on the boundary conditions.

For point constraints, see Kimmel-Yavneh Algebraic Multi-Grid approach.

),( yxfz

Page 13: ron Numerical Geometry of Images: Shape Reconstruction Ron Kimmel Geometric Image Processing Lab Computer Science Department Technion-Israel.

Multigrid Poisson solvers

Solve a Poisson equation

),( yxfz

2 2 2

solve ( , )

search for in a coarse scale 2

ˆsolving for ,

we have that our solution is given by

ˆ.

h h h

h h h h

h

h h h

z f x y

r z f

r h

z r

z r

z z z

Page 14: ron Numerical Geometry of Images: Shape Reconstruction Ron Kimmel Geometric Image Processing Lab Computer Science Department Technion-Israel.

Shape from shading

Horn: characteristic strip expansion Bruckstein: equal height contours Rouy-Turin: Minimal cost approach Kimmel-Sethian: Fast Marching

Page 15: ron Numerical Geometry of Images: Shape Reconstruction Ron Kimmel Geometric Image Processing Lab Computer Science Department Technion-Israel.

Summary: Shape reconstruction

Shape from shading Explicit methods (Horn/Bruckstein)

Page 16: ron Numerical Geometry of Images: Shape Reconstruction Ron Kimmel Geometric Image Processing Lab Computer Science Department Technion-Israel.

Open Questions

A working automatic shape from stereo algorithm does not exist (but we are getting there).

www.cs.technion.ac.il/~ron