You are on page 1of 9

Pattern Recognition 41 (2008) 3749 -- 3757

Contents lists available at ScienceDirect

Pattern Recognition
journal homepage: w w w . e l s e v i e r . c o m / l o c a t e / p r

Shape from shading using graph cuts


Ju Yong Chang a , Kyoung Mu Lee b,∗ , Sang Uk Lee b
a Mitsubishi Electric Research Laboratories, 201 Broadway, Cambridge, MA 02139, USA
b Department of Electrical Engineering, ASRI, Seoul National University, 599 Gwanangno, Gwanak-gu, Seoul 151-742, Republic of Korea

A R T I C L E I N F O A B S T R A C T

Article history: In this paper, we investigate the applicability of graph cuts to the SFS (shape-from-shading) problem.
Received 30 November 2005 We propose a new semi-global method for SFS using graph cuts. The new algorithm combines the local
Received in revised form 9 April 2008 method proposed by Lee and Rosenfeld [C.H. Lee, A. Rosenfeld, Improved methods of estimating shape
Accepted 19 May 2008
from shading using the light source coordinate system, Artif. Intell. 26 (1985) 125–143] and a global
method using an energy minimization technique. By employing a new global energy minimization formu-
Keywords: lation, the convex/concave ambiguity problem of Lee and Rosenfeld's method can be resolved efficiently.
Shape from shading A new combinatorial optimization technique, the graph cuts method, is used for the minimization of the
Graph cuts proposed energy functional. Experimental results on a variety of synthetic and real-world images show
that the proposed algorithm reconstructs the 3-D shape of objects very efficiently.
© 2008 Elsevier Ltd. All rights reserved.

1. Introduction Global approaches attempt to recover the entire surface by minimiz-


ing some energy functional associated with the surface to be esti-
Shape-from-shading (SFS) has been a central problem in the field mated. However, because the problem is ill posed, additional con-
of computer vision since the early years. This problem involves how straints are needed to solve it. Usually, the smoothness constraint
to compute the 3-D shape of an object from the brightness varia- such as a measure of “departure from smoothness” is added to the
tions in an intensity image of the object. To solve this problem, it is original energy functional. In general, if we assume that the surface
important to model how the images are formed. Although many re- is smooth, there is a unique solution to the image irradiance equa-
flectance models have been proposed, most SFS techniques assume tion [11]. This justifies the use of the smoothness constraint. This
the Lambertian reflection model. According to the Lambertian model, combined energy functional is commonly minimized by a variational
the brightness information at a pixel in an image depends only on method. Global techniques have been shown to be more generally
the albedo, surface normal vector, and the light source direction. This applicable to different types of input images and are more robust to
is illustrated in Fig. 1. Using this Lambertian model, the image for- noise than local techniques. Moreover, they also generate more ac-
mation process can be modeled by the following image irradiance curate results. However, they have demerits such as the tendency to
equation: oversmooth.
Local approaches involve the use of the local brightness informa-

→ − →
I(x, y) =  S · N (x, y). (1) tion of an image, and they recover surface patches that are subse-
quently quilted together. These local techniques are fast and recover
the local features of a surface more accurately. However, they are
Thus, the SFS problem can be defined by finding the best way to sensitive to noise and require some restrictive assumptions about
reconstruct the geometric information of the object that satisfies this the surface.
image irradiance equation. The reconstructed geometric information
can either be the surface depth or a set of surface normals, often
1.1. Summary of our results
described as a needle map.
A number of different approaches to solve the SFS problem have
In this paper, we propose a new semi-global SFS technique based
been proposed [15]. These methods can be divided into two broad
on graph cuts. This method minimizes a new energy functional in
categories: global approaches [1–8] and local approaches [9,10].
which the local and global methods are combined together. The pro-
∗ posed energy functional is naturally represented in a discrete formu-
Corresponding author. Tel.: +82 2 880 1743; fax: +82 2 878 1452.
E-mail addresses: jchang@merl.com (J.Y. Chang), kyoungmu@snu.ac.kr (K.M. Lee), lation. Thus, we can use a new combinatorial optimization technique,
sanguk@ipl.snu.ac.kr (S.U. Lee). the graph cuts method. This method is known to be able to find a

0031-3203/$ - see front matter © 2008 Elsevier Ltd. All rights reserved.
doi:10.1016/j.patcog.2008.05.020
3750 J.Y. Chang et al. / Pattern Recognition 41 (2008) 3749 -- 3757

A discrete approximation to the Laplacian operator


→2 4 −
→ −

∇ N ij  ( N ij − N ij ) (4)
2
was used to change the Euler equation into the following discrete
form:

→ −
− →− → 4 −→ −
→ −

(Iij −  N ij · S ) S + ( N ij − N ij ) − ij N ij = 0. (5)
2
−→
Here N ij is the average of the normals in the neighborhood around
the point (i, j).


Assuming a known light source direction, S , the iterative scheme


Fig. 1. S indicates the light source direction and −

n indicates the surface normal −

for N was developed by rearranging the above equation and taking
vector.
only the direction of the vector by dropping the constant term 1/(1+
ij (2 /4)). The final iterative equations are
local minimum closer to the global minimum than other min-


→(k) →(k) −
− → −

⎨−→ = N ij +  (Iij − N ij · S (k) ) S (k) ,
imization techniques, and it has relatively small computational (k+1) 2
m ij
demand [12]. 4 (6)
⎩−→(k+1) −
N ij =→
(k+1) −
m ij /→
(k+1)
m ij ,
This new technique offers a number of advantages. First, the local
features are well reconstructed because the new technique has the
nature of local SFS approaches. Second, the global minimization and where  is the distance between adjacent pixels in the image.
smoothing process make the new technique robust to noise. Third, The principal criticism to the Brooks and Horn's algorithm and
this new method also improves the data consistency of the recovered to similar global approaches [1,3] is the tendency to oversmooth
needle map, so it prevents oversmoothing of recovered surfaces. the recovered needle map. Specifically, in Eq. (6), the smoothness
Fourth, the computational demand in this method is less than those term dominates the data term. Because the smoothness constraint is
in other global SFS approaches. formulated in terms of the directional derivatives of the needle map,
it is trivially minimized by a flat surface. Thus, the conflict between
the data and the model leads to a strongly smoothed needle map
1.2. Organization of the paper and loss of detail.

The rest of the paper is organized as follows. In Section 2, we


briefly summarize the variational approach as the global SFS method 2.2. Local method
and also review the local SFS method proposed by Lee and Rosenfeld.
In Section 3, we formulate our energy functional and compare it Lee and Rosenfeld proposed a local SFS algorithm that determines
with existing functions. Section 4 contains our consideration of the the surface normal locally using both intensity and intensity deriva-
applicability of graph cuts to the proposed energy functional and tive information at a pixel in an image [9]. The surface normal can
shows how we use graph cuts to minimize the given functional. be represented as
Section 5 presents the experimental results for a variety of synthetic −

N = (Nx , Ny , Nz ) = (sin  cos , sin  sin , cos ), (7)
and real-world images. Finally, the conclusion of the paper is given
in Section 6. where  and  are the slant and tilt of the surface normal, respec-
tively. If we assume the light source-centered coordinates, then the
2. Existing SFS methods image irradiance equation can be rewritten as

→ − →
2.1. Global method I(x, y) =  S · N
= (0, 0, 1) · (sin  cos , sin  sin , cos )
Brooks and Horn [13] combined brightness constraint, smooth- =  cos . (8)
ness constraint, and unit normal constraint, and minimized the fol-
lowing energy functional: Hence, under the assumption that the albedo  is known, the slant 
  of the surface normal can be determined by the intensity information

→ − → −
→ −

{(I −  S · N )2 + ( N x 2 +  N y 2 ) at the pixel.

→ To determine the tilt  of the surface normal, Lee and Rosen-
+ ( N 2 − 1)} dx dy, (2) feld assumed that a local surface patch could be approximated by a

→ spherical patch. Under this assumption, they proved that the tilt of
where N is the surface normal,  is a scalar that weighs the relative the surface normal could be obtained by
importance of the smoothness term, and  is a Lagrangian multiplier.
The first term in the above equation is the brightness constraint that Iy cos S − Ix sin S
 = arctan , (9)
is derived from the image irradiance equation (1). The second term Ix cos S cos S + Iy cos S sin S
is the smoothness constraint that ensures a smooth surface in order
to stabilize the convergence to a unique solution. The third term is where Ix and Iy are the intensity derivatives along the x and y direc-
the unit normal constraint to enforce the recovered surface normals tions, and S and S are the slant and tilt of the light source direc-
to be unit vectors. tion, respectively. The slant and tilt of the surface normal computed
The minimization of the above functional was done through the by Eqs. (8) and (9) is a value in the light source-centered coordinates.
calculus of variation. We can show that the function given in Eq. (2) Thus, it should be transformed into the value in the viewer-centered
has the following Euler equation: coordinates to determine the final surface normal vector.
In determining the tilt of the surface normal using Lee and Rosen-

→ − →− → −
→ −

(I −  S · N ) S + ∇ 2 N −  N = 0. (3) feld's method, there exists an ambiguity in convex and concave cases.
J.Y. Chang et al. / Pattern Recognition 41 (2008) 3749 -- 3757 3751

The determination of the tilt value by Eq. (9) results from the as-
sumption that the surface is locally concave. Thus, the tilt in the
convex surface has an opposite direction to the tilt in the concave
case. Besides this convex/concave ambiguity, this method tends to
be more sensitive to noise as the distance between the viewer and
the light source increases.

3. Proposed energy function

As stated above, Lee and Rosenfeld's algorithm derives the surface


normal vectors by using intensity and intensity gradient informa-
tion. However, under a locally spherical assumption, there remains
the ambiguity such that which one we can select among two possi-
ble surface normals. To resolve this problem, we need to assume the
surface type (convex/concave) before applying the algorithm. How-
ever, in this case, we cannot avoid large errors in regions where the
assumption about the surface type is not correct. In this paper, we
aim to address this problem by proposing an algorithm that resolves
the ambiguity globally using energy minimization formulation.
Suppose that − →
n p (0) and −

n p (1) are the two allowable surface
normal vectors under the convex and concave assumptions at a pixel
p, respectively, as shown in Fig. 2(a). The convex/concave regular-
ization constraint can then be defined by

(X) = Vpq (Xp , Xq )
(p,q)∈N

= −

n p (Xp ) − −

n q (Xq ), X ∈ {0, 1}, (10)
(p,q)∈N

where N is the set of all neighboring pixels, Vpq is an interaction


function, and X is a function that is defined at each pixel, which
can take 0 or 1. By minimizing the above functional, we enforce
the surface normal to change gradually so that the convex/concave
ambiguity is resolved. This assumption is suitable for general 3-D
objects.
Note that the above energy function (10) is a special case of the
following general energy function:

() = Vpq (p , q )
(p,q)∈N

= −

n p (p ) − −

n q (q ), (11)
(p,q)∈N

where  is the tilt of the surface normal. The allowable surface normal
vector −
→n p (p ) is shown in Fig. 2(b). If we enforce the tilt of the above
energy function to have the convex case value or the concave case
value in the local method, we can obtain the energy function in Eq. Fig. 2. (a) Two allowable normal vectors − →
n p (0) and −
→n p (1), (b) allowable tilt p
and normal vector − →
n p (p ), and (c) allowable slant p , tilt p , and normal vector
(10). Furthermore, the above equation can be regarded as a special −
→n p (p , p ).
case of the following more general energy function:
 
(, ) = Dp (p , p ) + Vpq (p , q , p , q )
p (p,q)∈N In addition, we can think that Eq. (12) is the discrete form of the
 −
→ → energy function used in Brooks and Horn's method, that is Eq. (2).
= (I −  S · −
n (p , p ))2
If we deal with an object for which the local spherical assumption
p
 is not suitable, we can relax the constraint in Eq. (10) to permit
+ −

n p (p , p ) − −

n q (q , q ). (12) other tilt values besides the two kinds of tilt values determined by
(p,q)∈N the local method. To do so, we add the following term to the energy
function in Eq. (11):
The allowable surface normal vector − →n p (p , p ) is shown in Fig. 2(c).

If we assume that the image irradiance equation is always guaran- Rp (p ), (14)

→ →
teed, then (I −  S · −
n (p , p )) = 0 is satisfied at each pixel p; there- p
fore, the above equation is transformed into Eq. (11). In this case, the
slant of the normal vector can be obtained from Eq. (8) such that and the resulting energy function is as follows:
   
 = arccos
I(x, y)
. (13)
() = Rp (p ) + Vpq (p , q ). (15)
 p (p,q)∈N
3752 J.Y. Chang et al. / Pattern Recognition 41 (2008) 3749 -- 3757

In this equation, Rp , as a function defined at each pixel p, assigns a


large penalty if the difference between the tilt at pixel p and the two
tilt values determined by the local method is large. If we assign 0
to Rp when the tilt value at pixel p is the tilt in convex or concave
cases and assign ∞ to Rp otherwise, then Eq. (15) is transformed
into Eq. (10). In this paper, we consider the above energy functions
(10) and (15).
The proposed energy function has several advantages over those
used in conventional global approaches. First, the proposed energy
function does not include the brightness constraint, while most con-
ventional energy functions involve this constraint as well as the
smoothness constraint. The brightness constraint is derived directly
from the image irradiance equation (1) and indicates the total bright-
ness error of the reconstructed image compared with the input im-
age. Therefore, conventional global methods do not guarantee that
the reconstructed pixel intensity is identical to the real pixel in-
tensity. Hence, the reconstructed surface normal vector may have a
Fig. 3. The regularity condition is not satisfied at these pixels p and q. slant value that differs from the slant value derived from Eq. (13). In
comparison, in the case of the energy function proposed in this pa-
per, the allowable surface normal vectors at each pixel have the slant
value determined by the image intensity. This results in an imposing
brightness constraint as a hard constraint. Actually, the normal in-
tegration process for obtaining the final surface alters the slant of a
surface normal, but our energy formulation reduces the oversmooth-
ing effect of the recovered surface. Similarly, Worthington recently
proposed a geometric framework for the same purpose [7]. Second,
in contrast with most conventional energy functions that allow an
uncountable number of surface normal vectors as candidates, in the
proposed energy function, the number of allowable surface normal
vectors at each pixel is limited by two, convex and concave cases.
Various continuous or discrete optimization methods can be used
for minimizing the proposed energy function. In this work, we use
the graph cuts algorithm because it is known as one of the most
powerful optimization techniques in computer vision.


→ −

Fig. 4. New difference vector d , original difference vector −

a , bias vector b . 4. Energy minimization using graph cuts

4.1. Graph cuts method

To minimize the proposed energy function, we use the graph cuts


method that was introduced as a new combinatorial energy mini-
mization technique [12]. The basic technique of graph cuts is to con-
struct a specialized graph for the energy function to be minimized,
such that the minimum cut on the graph also minimizes the energy
(either globally or locally). The minimum cut can in turn be com-
puted very efficiently by max flow algorithms. This method has been
successfully used for a wide variety of vision problems. The output
of this algorithm is generally a solution with some interesting the-
oretical quality guarantee. In some cases, it is the global minimum,
and in other cases, a local minimum in a strong sense that is within
a known factor of the global minimum. The experimental results
produced by graph cuts are also quite good and the computational
demand is relatively small.
Minimizing an energy function via graph cuts, however, re-
mains a technically difficult problem. We need to construct ap-
propriate graph for a given energy function, and in some cases,
the construction becomes fairly complex. Recently, Kolmogorov
and Zabih characterized the class of energy functions that can
be minimized via graph cuts and proposed a general-purpose
graph model that minimizes any energy function in this class [14].
Some of the main results in their paper are as follows. Roughly
speaking, let a function E of n binary variables be called graph-
representable if we can find the global minimum of the function E
Fig. 5. The depth map for the synthetic image, Mozart.
via graph cuts. According to their paper, E is graph-representable
if and only if E is regular. Let E be a function having the following
J.Y. Chang et al. / Pattern Recognition 41 (2008) 3749 -- 3757 3753

Fig. 6. Synthetic and real-world images for experiments: (a) Mozart image, (b) Lenna image, (c) Pepper image, and (d) Vase image.

form: Thus, to use graph cuts, we need to modify the original energy
  function properly. Among several graph cuts algorithms, we employ
E(x1 , . . . , xn ) = Ei (xi ) + Ei,j (xi , xj ). (16)
the expansion move algorithm, which was introduced in Ref. [12].
i i<j
Now, let us consider the simple energy function in Eq. (10). This
E is then called regular if it satisfies the following inequality: function can be expressed in the following form:
 
Ei,j (0, 0) + Ei,j (1, 1)  Ei,j (0, 1) + Ei,j (1, 0). (17) () = Rp (p ) + Vpq (p , q ),
p (p,q)∈N
Regularity is thus an extremely important property as it allows en-

ergy functions to be minimized by graph cuts. Moreover, without the 0 if p is a tilt in convex or concave case,
regularity constraint, the problem becomes intractable. Their second Rp (p ) =
result shows that minimizing an arbitrary non-regular function is ∞ otherwise,
NP-hard. Vpq (p , q ) = −

n p (p ) − −

n q (q ). (18)

4.2. SFS using graph cuts As we have already seen, this function is a special case of Eq. (15).
According to Ref. [12], the expansion move algorithm can be used
In this section, we investigate the applicability of graph cuts to the when Vpq satisfies the metric condition as follows:
proposed energy function. As stated above, to apply the graph cuts
method to the energy minimization problem, the energy function Vpq (, ) = 0 ⇐⇒  = ,
has to satisfy the regularity condition. First, let us investigate the
Vpq (, ) = Vpq ( , )  0,
simple energy function in Eq. (10). Note that, however, the proposed
energy function in Eq. (10) may not satisfy the regularity condition Vpq (, )  Vpq (,
) + Vpq (
, ). (19)
at some pixels. This is illustrated in Fig. 3. Thus, we can conclude
that we cannot find the global minimum of this energy function However, Vpq in Eq. (18) does not satisfy the metric condition since
via graph cuts. Moreover, because this simple energy function is a Vpq (k, k) = −

n p (k) − −

n q (k) = 0. This is because the slant values at
special case of more general energy functions like Eqs. (11), (12), and pixels p and q are different. Now, let us modify Vpq as
(15), we can conclude that the general energy function for the SFS
problem cannot be minimized globally via graph cuts. Furthermore, Vpq (p , q , p , q ) = −

n (, p ) − −

n (, q ),
according to the NP-hardness theorem, it is very likely that the global
minimization of these energy functions is NP-hard.  = min(p , q ). (20)
3754 J.Y. Chang et al. / Pattern Recognition 41 (2008) 3749 -- 3757

Fig. 7. Surface reconstruction results for a synthetic image of Mozart: (a) surface reconstructed using Horn and Brooks method, (b) surface reconstructed using Tsai and
Shah method, (c) surface reconstructed using Lee and Rosenfeld method, and (d) surface reconstructed using the proposed method.

This new Vpq function is the magnitude of a new difference vec- 5. Experiments
tor that is obtained by subtracting the original difference vector to
the bias vector. This is illustrated in Fig. 4. Note that this new func- In the experiments, we have used one synthetic image and three
tion satisfies the metric condition. Now, we can apply the expansion real images. The ground truth depth map for the synthetic image,
move algorithm to this new energy function. Furthermore, the gen- Mozart, is shown in Fig. 5. The synthetic image generated from this
eral energy function (15) that has new Vpq can also be minimized surface with light source direction (1, 0, 1) is shown in Fig. 6(a). The
by the expansion move algorithm. light source directions of the three real images as shown in Figs. 6(b),
J.Y. Chang et al. / Pattern Recognition 41 (2008) 3749 -- 3757 3755

Fig. 8. Surface reconstruction results for a real image of Lenna: (a) surface reconstructed using Horn and Brooks method, (b) surface reconstructed using Tsai and Shah
method, (c) surface reconstructed using Lee and Rosenfeld method, and (d) surface reconstructed using the proposed method.

6(c) and 6(d) were estimated by Lee and Rosenfeld's method [9]. rithms including one global SFS algorithm of Brooks and Horn [13]
The estimated light source directions of the Lenna, Pepper, and Vase and two local SFS algorithms of Lee and Rosenfeld [9] and Tsai and
images were (1. 5, 0. 866, 1), (0. 766, 0. 642, 1), and (−0. 939, 1. 867, 1). Shah [10]. Brooks and Horn's approach is usually initialized by esti-
We have performed several experiments on the above images mating the occluding boundary normals with all other normals set
using our algorithm. The energy function (15) has been used for all to point in the light source direction. However, for fair comparison,
experiments. We have also compared this with existing SFS algo- we initialized Brooks and Horn's algorithm with all normals set to
3756 J.Y. Chang et al. / Pattern Recognition 41 (2008) 3749 -- 3757

Fig. 9. Close-up results for a synthetic image of Mozart: (a) original surface of Mozart with a red rectangle indicating the close-up region. (b) Surface reconstructed using
Lee and Rosenfeld method and (c) surface reconstructed using the proposed method.

Fig. 10. Surface reconstruction results for various real images: (a) surface reconstructed from Pepper image and (b) surface reconstructed from Vase image.

point in the light source direction. Moreover, in Lee and Rosenfeld's was necessary. To integrate the normal map, we used Frankot and
algorithm, we assumed that the surface of the object is convex. Chellappa's method [16] to obtain the modified integrable normal
For the case of Tsai and Shah's algorithm, the depth map was di- maps, while the resultant surfaces were obtained by applying simple
rectly computed. In contrast, in Brooks and Horn, Lee and Rosenfeld, path integration to the normal map.
and our SFS algorithms, output is the estimated normal map. There- The reconstruction results for the synthetic image, Mozart, and
fore, to generate the final surfaces, the normal map integration step the real image, Lenna, are shown in Figs. 7 and 8, respectively. First,
J.Y. Chang et al. / Pattern Recognition 41 (2008) 3749 -- 3757 3757

Figs. 7(a) and 8(a) show the reconstructed surface obtained by Brooks Acknowledgments
and Horn's algorithm. We can see that the surfaces reconstructed by
this algorithm lost details due to oversmoothing. Second, Figs. 7(b) This work was supported in part by the ITRC program by Ministry
and 8(b) show the result of Tsai and Shah's algorithm. We can see of Information and Communication and in part by Defense Acquisi-
that their method works well on some smooth regions. However, it tion Program Administration and Agency for Defense Development,
produces noisy results in regions where there are abrupt changes Korea, through the Image Information Research Center under the
of intensities. Moreover, in the result of the Mozart image, it seems contract UD070007AD.
that the global convex/concave ambiguity has not been resolved.
Third, Figs. 7(c) and 8(c) show the result of Lee and Rosenfeld's al- References
gorithm. In this case, we can see that this algorithm produced noisy
surfaces despite its relatively good recovery of local features. Usu- [1] K. Ikeuchi, B.K.P. Horn, Numerical shape from shading and occluding boundaries,
Artif. Intell. 17 (3) (1981) 141–184.
ally, this method becomes more sensitive to noise as the distance [2] B.K.P. Horn, Height and gradient from shading, Int. J. Comput. Vision (1989)
between the viewer and the light source increases. Finally, Figs. 7(d) 37–75.
and 8(d) show the result of the proposed algorithm. The proposed al- [3] R. Szeliski, Fast shape from shading, Comput. Vision Graphics Image Process.:
Image Understanding 53 (1991) 129–153.
gorithm captures sufficient surface details while performing proper [4] K.M. Lee, C.C.J. Kuo, Shape from shading with a linear triangular element surface
smoothing process, thus recovering more suitable surfaces. In con- model, IEEE Trans. Pattern Anal. Mach. Intell. 15 (8) (1993) 815–822.
trast to Lee and Rosenfeld's algorithm, the proposed method can re- [5] J. Oliensis, P. Dupuis, A global algorithm for shape from shading, in: Proceedings
of the International Conference on Computer Vision, 1993, pp. 692–701.
solve the convex/concave ambiguity by using the energy minimiza- [6] R. Kimmel, A.M. Bruckstein, Global shape from shading, Comput. Vision Image
tion framework. Furthermore, the result improvement can be seen Understanding 62 (3) (1995) 360–369.
in the close-up image shown in Fig. 9. Fig. 10 shows the result of our [7] P.L. Worthington, E.R. Hancock, New constraints on data-closeness and needle-
map consistency for shape-from-shading, IEEE Trans. Pattern Anal. Mach. Intell.
algorithm on additional real-world images, Pepper and Vase. These 21 (12) (1999) 1250–1267.
results demonstrate that the proposed algorithm is able to recon- [8] Q. Zhu, J. Shi, Shape from shading: recognizing the mountains through a global
struct the 3-D surface of various real-world objects successfully. view, in: Proceedings on Computer Vision and Pattern Recognition, 2006, pp.
1839–1846.
[9] C.H. Lee, A. Rosenfeld, Improved methods of estimating shape from shading
6. Conclusions using the light source coordinate system, Artif. Intell. 26 (1985) 125–143.
[10] P.S. Tsai, M. Shah, Shape from shading using linear approximation, Image Vision
In this paper, we proposed a new semi-global SFS algorithm that Comput. 12 (8) (1994) 487–498.
[11] J. Kain, D.N. Ostrov, Numerical shape-from-shading for discontinuous
combines local and global energy minimization frameworks. The photographic images, Int. J. Comput. Vision 44 (3) (2001) 163–173.
method uses the surface normal vectors obtained by Lee and Rosen- [12] Y. Boykov, O. Veksler, R. Zabih, Fast approximate energy minimization via graph
feld's method as the allowable normals at each pixel. By employing a cuts, IEEE Trans. Pattern Anal. Mach. Intell. 23 (11) (2001) 1222–1239.
[13] M.J. Brooks, B.K.P. Horn, Shape and source from shading, in: Proceedings of the
new energy minimization formulation based on the convex/concave International Joint Conference on Artificial Intelligence, 1985, pp. 932–936.
regularization constraint of local normal vectors, the method deter- [14] V. Kolmogorov, R. Zabih, What energy functions can be minimized via graph
mines the final needle map. To minimize the proposed energy func- cuts?, IEEE Trans. Pattern Anal. Mach. Intell. 26 (2) (2004) 147–159.
[15] R. Zhang, P.S. Tsai, J.E. Cryer, M. Shah, Shape from shading: a survey, IEEE
tion, we applied the graph cuts technique. The results on synthetic Trans. Pattern Anal. Mach. Intell. 21 (8) (1999) 690–706.
and real world objects reveal that the proposed algorithm produces [16] R.T. Frankot, R. Chellappa, A method for enforcing integrability in shape from
more robust and enhanced results than local methods, and it exhibits shading, IEEE Trans. Pattern Anal. Mach. Intell. 10 (4) (1988) 439–451.
local features better than other global methods.

About the Author—JU YONG CHANG received the B.S. and Ph.D. degrees in Electrical Engineering and Computer Science from Seoul National University, Seoul, Korea, in
2001 and 2008, respectively. In February 2008, he joined the Mitsubishi Electric Research Laboratories at Cambridge, MA, as a postdoctoral researcher. His current research
interests are 3-D shape modeling and scene reconstruction.

About the Author—KYOUNG MU LEE received the B.S. and M.S. degrees in Control and Instrumentation Engineering from Seoul National University, Seoul, Korea, in 1984
and 1986, respectively, and Ph.D. degree in Electrical Engineering from the University of Southern California, Los Angeles, in 1993. From 1993 to 1994, he was a research
associate in the Signal and Image Processing Institute at the University of Southern California. He was with the Samsung Electronics Co. Ltd. at Suwon in Korea as a senior
researcher from 1994 to 1995, where he worked on developing industrial real-time vision systems. From 1995 to 2003, he was an Assistant and an Associate Professor in
the Department of Electronics and Electrical Engineering of the Hong-Ik University in Seoul, Korea. In September 2003, he joined the School of Electrical Engineering and
Computer Science at Seoul National University, where he is now a Professor. His current primary research interests include computational vision, shape from X, 2-D and
3-D object recognition, human-computer interface, and visual navigation.

About the Author—SANG UK LEE received the B.S. degree from Seoul National University, Seoul, Korea, in 1973, the M.S. degree from Iowa State University, Ames, in
1976, and Ph.D. degree from the University of Southern California, Los Angeles, in 1980, all in electrical engineering. From 1980 to 1981, he was with the General Electric
Company, Lynchburg, VA, working on the development of digital mobile radio. From 1981 to 1983, he was a Member of Technical Staff, M/A-COM Research Center, Rockville,
MD. In 1983, he joined the Department of Control and Instrumentation Engineering at Seoul National University as an Assistant Professor, where he served as the Head of
the School of Electrical Engineering and Computer Science. He is also affiliated with the Automation and Systems Research Institute and the Institute of New Media and
Communications at Seoul National University. He was the President of the Korean Institute of Communication Science in 2005. His current research interests are in the
areas of image and video signal processing, digital communication, and computer vision.
Dr. Lee served as an Editor-in-Chief for the Transaction of the Korean Institute of Communication Science from 1994 to 1996. He was an Associate Editor for IEEE Transactions
on Circuits and Systems for Video Technology from 2002 to 2005, and was on the Editorial Board of the EURASIP Journal of Applied Signal Processing from 2003 to 2004.
He is currently on the Editorial Board of the Journal of Visual Communication and Image Representation. He is a member of Phi Kappa Phi.

You might also like