1 Fast Marching Method for generic Shape From Shading E. Prados & S. Soatto RFIA 2006 January 2006, Tours, France

1

Fast Marching Method for generic Shape From Shading

E. Prados & S. Soatto

RFIA 2006January 2006, Tours, France

2

Hypotheses: Lambertian Reflectance, Lighting:

Single punctual light source, Located at the infinity or optical center, Possibility to take into account the

attenuation of the light due to the distance, Orthographic or perspective camera .

Explicit generic SFS equation

3


Particular case of the generic PDE:

F(u)

[Prados:PhD'04]

+

4


Particular case of the generic PDE:

Particular characteristics: • Dependency in u (no only in ∇u).• Solution not necessarily increasing along the characteristic curves...

Current Fast Marching Methods do not apply.

F(u)+

5

Iterative method FMMIterative method

Fast Marching Method (FMM)

• One-pass method for solving numerically PDEs≠ iterative methods: – No threshold as stopping criterion– Optimal number of updates → Low computation time– Based on front propagation.

6




7




8




9




• Many applications: – Path planning [Kimmel-Sethian:01]

– Geometric optics [Wenwang:03]

– Image processing and computer vision [L.Cohen:05]

– Extensive list, see [Sethian:99].

10

(x,y-1)

(x+1,y)(x,y)(x-1,y)

(x,y+1)


• Basic method [Sethian:99, Dijkstra:59] |∇u|= g(x), (Eikonal equation) 4-neighborhood scheme.

• Most recent extension: OUM [Sethian-Vladimirsky:03]

supa {– f(x,a) a.∇u -1} = 0, (2)

Scheme requiring very large neighborhood

(size depending on the anisotropy)

• Our extension (new algorithm): λ F(u) + H(x,∇u) = 0, (3)

where F is strictly increasing and H is convex 4-neighborhood scheme.

11

Our New Extension of the FMM

1. A New Numerical Scheme…

2. A New Causality…

12

A New Numerical Scheme• Preliminary step:

Legendre Tranform equation as a supremum

λ F(u) + H(x,∇u) = 0

λ F(u) + sup { – f(x,a).∇u(x) - l(x,a) } = 0

Cost function

13

A New Numerical Scheme

We approximate

And we choose the simplex, i.e. the si such that the

scheme is increasing with respect to t (represent u)

We then choose si = si(x,a) := sign fi(x,a) ,

In other words, we choose the simplex which contains the optimal trajectory (see next section…).

λ F(t) +

t u(x)

si ∈ {+1, -1}

14

A New Numerical Scheme

• Our scheme is provably consistent and monotonic

• Only direct neighborhoods in SFS with regular mesh: 4-neighborhood.

• Equation depends on u

15

A New Causality and Reinterpretation

Key point: Distinction of the causality and of the simultaneous integration.

1. Causality: Theoretic propagation of the information; ensures that we compute approximation of the viscosity solution !

The information propagation follows specific curves:

the optimal trajectories:

The solution can be computed by direct integration along these curves:

Trying to compute the solution curve after curve is numerically unstable → Simultaneous integration along all the curves…

16

Simultaneous integration: ensures the numerical stability !

→ Propagation front ;

How to choose the propagation front? Many propagation fronts following the optimal trajectories can be designed!

17




18




19




How to define the propagation front? idea: introduction of a “cost” C such that level sets of C correspond with propagated front.

20

Simultaneous integration:

Level sets of u

Optimal trajectory

→ All previous SFS methods choose C = u ( u is the solution)

inconsistent with the causality when the solution does not increase along the optimal trajectories !

→ We show how to define an appropriate cost C - which is always consistent with optimal trajectories, - which allows to define and compute simply and practically

the update order…

→ This cost C is based on the notion of subsolution: ψ …

21


Values computed using the new scheme

Domain where we know the solution

Choice of the pixel based on causality

22

Algorithm F = points éloignésA = points acceptés

C = points considérés

23

A Practical and Detailed Example:The classical Rouy/Tourin equation

Associated Modeling:• Lambertian and homogeneous reflectance,

• Single oblique and far light source,

• orthographic projection.

Resulting PDE:

where I(x) is the image.

Associated cost function:

have arbitrary signs…

Solution does not increase along the characteristic curves,All previous Fast Marching Methods do not directly apply !

Associated subsolution:

24

Focus on the improvement due to the new Causality

original image

groundtruth

Reconstruction with thenew and correct causality

oblique view

profile viewprofile view

oblique view

Computed solution

groundtruth

Reconstruction with the classical causality

25

SFS with attenuation of light due to distance

Associated Modeling:• Lambertian and homogeneous reflectance,

• Single proximal light source (optical center),

• perspective projection.

Resulting PDE: I(x) = image and f = focal length.

and

All previous Fast Marching Methods do not directly apply

image reconstruction

26

References

[Kimmel-Sethian:01] R. Kimmel and J.A. Sethian. Optimal algorithm for shape from shading and path planning. JMIV, 14(2):237–244, May 2001.

[L.Cohen:05] L. Cohen. Minimal paths and fast marching methods for image analysis. In Mathematical Models in Computer Vision: The Handbook, Springer, 2005.

[Sethian:99] J.A. Sethian. Level Set Methods and Fast Marching Methods. Cambridge University Press,1999.

[Sethian-Vladimirsky:03] J.A. Sethian and A. Vladimirsky. Ordered upwind methods for Hamilton–Jacobi equations:Theory and algorithms. SIAM J. on Num. Ana. 41(1), 2003

[Prados:04] E. Prados. Application of the theory of the viscosity solutions to the Shape From Shading problem. PhD thesis, 2004.

References

Documents

1 Fast Marching Method for generic Shape From Shading E. Prados & S. Soatto RFIA 2006 January 2006, Tours, France