Smooth Shape-Aware Functions with Controlled Extrema...Smooth Shape-Aware Functions with Controlled Extrema Alec Jacobson1 Tino Weinkauf2 Olga Sorkine1 August 9, 2012 1ETH Zurich 2MPI

Smooth Shape-Aware Functions with Controlled Extrema Alec Jacobson1

Tino Weinkauf2

Olga Sorkine1

August 9, 2012

1ETH Zurich 2MPI Saarbrücken

#

Real-time deformation relies on smooth, shape-aware functions

August 9, 2012 Alec Jacobson 2

input shape + handles

#



precompute weight functions

#



deform handles à deform shape

#



#



#

unconstrained [Botsch & Kobbelt 2004]


local max

local min

Spurious extrema cause distracting artifacts

#

unconstrained [Botsch & Kobbelt 2004]


local max

local min

Spurious extrema cause distracting artifacts

#

bounded [Jacobson et al. 2011]


local max

local min

Bounds help, but don’t solve problem

#



local max

local min

Bounds help, but don’t solve problem

#



local max

local min

Gets worse with higher-order smoothness

oscillate too much

#



local max

local min

Gets worse with higher-order smoothness

oscillate too much

#

our


local max

local min

We explicitly prohibit spurious extrema

#

our


local max

local min

We explicitly prohibit spurious extrema

# August 9, 2012 Alec Jacobson 15

Same functions used for color interpolation




unconstrained [Finch et al. 2011]


Image courtesy Mark Finch







Our

#

Want same control when smoothing data


#



Exact, but sharp geodesic

#



Exact, but sharp geodesic

#



Exact, but sharp geodesic Smooth, but extrema are lost

#



Exact, but sharp geodesic Smooth and maintain extrema

#

Ideal discrete problem is intractable


Interpolation functions:

argmin

fE(f)

s.t. fmax

= gmax

fmin

= gmin

fj > fmax

fj < fmin

fi > min

j2N (i)fj

fi < max

j2N (i)fj

#



Data smoothing:

argmin

fE(f)

s.t. fmax

= gmax

fmin

= gmin

fj > fmax

fj < fmin

fi > min

j2N (i)fj

fi < max

j2N (i)fj

#



argmin

fE(f)

s.t. fmax

= known

fmin

= known

fj < fmax

fj > fmin

fi > min

j2N (i)fj

fi < max

j2N (i)fj

#



argmin

fE(f)

s.t. fmax

= known

fmin

= known

fj < fmax

fj > fmin

fi > min

j2N (i)fj

fi < max

j2N (i)fj

#

linear



fj

fmax

argmin

fE(f)

s.t. fmax

= known

fmin

= known

fj < fmax

fj > fmin

fi > min

j2N (i)fj

fi < max

j2N (i)fj

#

nonlinear

linear



fi

fj

argmin

fE(f)

s.t. fmax

= known

fmin

= known

fj < fmax

fj > fmin

fi > min

j2N (i)fj

fi < max

j2N (i)fj

#

nonlinear

linear

Assume we have a feasible solution


interior

handles

u“Representative function”

uj < umax

uj > umin

ui > min

j2N (i)uj

ui < max

j2N (i)uj

#

interior

handles

Assume we have a feasible solution


“Representative function” u

uj < umax

uj > umin

ui > min

j2N (i)uj

ui < max

j2N (i)uj

#

linear

Copy “monotonicity” of representative


At least one edge in either direction per vertex

argminf

E(f)

s.t. fmax

= known

fmin

= known

(fi � fj)(ui � uj) > 0 8(i, j) 2 E

#

Rewrite as conic optimization


Optimize with MOSEK

QP

Conic

#

We always have harmonic representative


argminu

1

2

Z

⌦

kruk2dV

s.t. umax

= 1

s.t. umin

= 0

#



argminu

1

2

Z

⌦

kruk2dV

s.t. umax

= 1

s.t. umin

= 0

#



argminu

1

2

Z

⌦

kruk2dV

s.t. umax

= 1

s.t. umin

= 0

#



argminu

1

2

Z

⌦

kruk2dV

s.t. umax

= 1

s.t. umin

= 0

Works well when no input function exists

#

Data energy may fight harmonic representative


Anisotropic input data

#



Anisotropic input data Harmonic representative

#




#




#



Anisotropic input data Resulting solution with large

#

If data exists, copy topology, too


Anisotropic input data [Weinkauf et al. 2010] representative

#

If data exists, copy topology, too


Anisotropic input data Resulting solution with large

#

Final algorithm is simple and efficient


●  Data smoothing: topology-aware representative §  Morse-smale + linear solve ~milliseconds

#




●  Interpolation: harmonic representative §  Linear solve ~milliseconds

#





●  Conic optimization §  2D ~milliseconds, 3D ~seconds

#





●  Conic optimization §  2D ~milliseconds, 3D ~seconds

Interpolation: functions are precomputed


We preserve troublesome appendages Bounded Our



#


#

Our weights attach appendages to body


Our method [Botsch & Kobbelt 2004, Jacobson et al. 2011]

#

Extrema glue appendages to far-away handles


[Botsch & Kobbelt 2004, Jacobson et al. 2011]

#

Extrema glue appendages to far-away handles


[Botsch & Kobbelt 2004, Jacobson et al. 2011]

#



Our method

#



Our method

#

Extrema distort small features


Unconstrained [Botsch & Kobbelt 2004]

weight of middle point

#



Unconstrained [Botsch & Kobbelt 2004]


#



Bounded [Jacobson et al. 2011]


#

“Monotonicity” helps preserve small features


Bounded [Jacobson et al. 2011]

Our

#

Spurious extrema are unstable, may “flip”


slightly larger region

#



slightly larger region

#


Unconstrained [Botsch & Kobbelt, 2004]

#



#



#


Bounded

#


Bounded

#

Lack of extrema leads to more stability

Our

#

Lack of extrema leads to more stability

Our

#

Even control continuity at extrema


Original

Direct extension of [Botsch & Kobbelt 2004]

[Botsch & Kobbelt 2004] + data term

Our method without data term

Our method with data term

#



Original





Original





#



Original





Original





#



Original





Original





#



Original





Original





#

Reproduces results of Weinkauf et al. 2010…


Original noisy data

#



Original noisy data

#



Original noisy data Simplified and smoothed

#



Original noisy data Simplified and smoothed

#

… but 1000 times faster

Alec Jacobson 80

30K vertices 5 seconds per solve

#


Alec Jacobson 81


#


Alec Jacobson 82


#

●  Copy “monotonicity” of harmonic functions

●  Reduces search-space, but optimization is tractable


Conclusion: Important to control extrema

#

●  Larger, but still tractable subspace? §  Consider all valid harmonic functions?


Future work and discussion

#

●  Larger, but still tractable subspace? §  Consider all valid harmonic functions?

●  Continuous formulation?


Future work and discussion

#

We thank Kenshi Takayama for his valuable feedback. This work was supported in part by an SNF award 200021_137879 and by a gift from Adobe Systems.


Acknowledgements

Alec Jacobson ([email protected])

Tino Weinkauf

Olga Sorkine

Smooth Shape-Aware Functions with Controlled Extrema

MATLAB Demo: http://igl.ethz.ch/projects/monotonic/

Smooth Shape-Aware Functions with Controlled Extrema...Smooth Shape-Aware Functions with Controlled Extrema Alec Jacobson1 Tino Weinkauf2 Olga Sorkine1 August 9, 2012 1ETH Zurich 2MPI

Documents