Professional Documents
Culture Documents
In this paper, we propose a new method for passive depth estimation based on the combination
of a camera with longitudinal chromatic aberration and an original depth from defocus (DFD)
algorithm. Indeed a chromatic lens, combined with an RGB sensor, produces three images with
spectrally variable in-focus planes, which eases the task of depth extraction with DFD. We first
propose an original DFD algorithm dedicated to color images having spectrally varying defocus
blurs. Then we describe the design of a prototype chromatic camera so as to evaluate experimentally
the effectiveness of the proposed approach for depth estimation. We provide comparisons with results
of an active ranging sensor and real indoor/outdoor scene reconstructions. © 2013 Optical Society of
America
OCIS codes: (110.0110) Imaging systems; (110.1758) Computational imaging; (100.0100) Image
processing; (100.3190) Inverse problems.
http://dx.doi.org/10.1364/AO.52.007152
1 1 1
ε Ds − − ; (1)
f d s
C. Grayscale Scene
In the case of a grayscale scene acquired with a chro-
matic camera, each RGB image actually originates
from the same scene so xR xG xB x. Thus the
observation model reduces to
2 3 2 3
yR H R d
6 7 6 7
Y 4 yG 5 4 H G d 5x N H gC x N: (5)
Fig. 4. Generic flowchart of the DFD algorithm for one image
yB H B d
patch.
1 Y T Qα;d;σ 2
n Y
pYjH gC d; σ 2n ; α ∝ jQα; d; σ 2n j2 e− 2 ; (8) where
2 3
with jAj corresponding to the product of the nonzero p1 −1
p −1
p
3 2 6
eigenvalues of a matrix A and 6 7
6 −1 7
T 6 p13 p
1
2
p
67
; (14)
Qα;d;σ 2n 4 5
p1 0 p2
3 6
1
2 I N;N − H gC dH gC dT H gC d αDT D−1 H gC dT ;
σn ⊗ stands for the Kronecker product. The observation
(9) model then writes
25 25
for depths varying from 1.3 to 3.5 m for five values
0 0 of μ. The value of μ that gives the lowest error metric
1.2 2.4 3.6 1.2 2.4 3.6 and standard deviation is 0.04. This result is consis-
Depth in m Depth in m tent with the value empirically chosen in [31]. This
75 75 value is fixed for all the other experiments of
the paper.
Std (cm)
Std (cm)
50 50
C. Chromatic Versus Achromatic Lens Depth Estimation
25 25 Performance Study
In this section, we use the CC-DFD algorithm on
0 0
1.2 2.4 3.6 1.2 2.4 3.6 simulated images for two imaging systems having
Depth in m Depth in m either a chromatic or an achromatic lens. The chro-
(a) (b) matic lens is identical to the one simulated in
Section 3.A. The achromatic lens has the same focal
CC−DFD algorithm length and f-number, i.e., respectively, 25 mm and 4,
GC−DFD algorithm but has a single in-focus plane that is put at 1.5 m in
order to have no depth estimation ambiguity nor
Fig. 5. Comparison of CC-DFD and GC-DFD algorithms using dead zone in the range of 2 to 3 m. The set of potential
error metric and standard deviation (std) computed over a collec- PSF triplets for each lens is built for depth varying
tion of (a) grayscale scenes and (b) color scenes. from 2 to 4 m with a step of 5 cm. For each imaging
system, we generate 120 image patches of size
20 × 20 pixels using scenes patches extracted from
produce each channel image. For the CC-DFD
natural color scenes presented in Fig. 5(b). For each
algorithm the value of μ is fixed at 0.04, a value
depth and each imaging system, the RGB images are
empirically determined in [31]. The depth estimation
obtained by convolution of scene patches with the
results for both algorithms and both sets are shown
corresponding PSFs triplet. White Gaussian noise
in Fig. 5. is added to the result with a standard deviation of
Figure 5(a) shows that for the set of grayscale 0.05, given that the scenes have a normalized inten-
scenes either the GC-DFD or the CC-DFD algorithm sity. Figure 6 shows the error metric and the stan-
leads to an error metric and a standard deviation dard deviation of the depth estimation obtained
under 10 cm, a value that is close to the depth sam- with the CC-DFD algorithm. The error metrics of
pling step of 5 cm used in the PSF set construction. both imaging systems are both low; however, the
Despite the fact that it is a grayscale scene, the
CC-DFD algorithm gives results very close to
20
Error metric in cm
100 AL
justifies the need for the algorithm based on the
modelization of the correlation between the RGB 50
channels, such as the proposed CC-DFD.
0
B. Setting of the Parameter μ 2 2.5 3 3.5
The depth estimation tests conducted in Section 3.A Depth in m
on the CC-DFD algorithm with color scenes are Fig. 6. Comparison of CC-DFD algorithm using either an achro-
reproduced for different values of μ. Table 1 shows matic lens (AL) or a chromatic lens (CL). Lens parameters for both
the mean error metric and mean standard deviation systems are given in Section 3.C.
R G B
1
1 2 3 4
5
3
with the CC-DFD algorithm can be compared to the
2 depth maps given by the Kinect. Homogeneous re-
gions, where depth can not be estimated, appear in
1
1 2 3 4 black in the depth map.
Fig. 13. Evaluation of depth estimation accuracy on real fronto- In the first two examples, the estimated depth
parallel grayscale scenes. Axes are in m. maps are noisier but consistent with the Kinect
depth map. The third example is a case in which
the Kinect depth map locally shows outliers. This
by the Kinect. Indeed, the Kinect accuracy ranges is due to occultations and thin surfaces where the
from a few millimeters at 1 m to 3 cm at 3 m [38], Kinect pattern does not reflect properly. In contrast,
which is lower than the sampling of depths in our po- the proposed chromatic DFD camera correctly esti-
tential set. For each scene, bias is comparable to the mates a depth map for such a complex scene.
PSF calibration step (5 cm) and standard deviation
varies from 3 to 10 cm between 1 and 3.5 m. The 2. Outdoor Situations
same experiment was repeated with the same tar-
gets printed in grayscale, leading to very similar In contrast with the Kinect that is sensible to the sun
results, presented in Fig. 13. These results demon- IR illumination, the proposed chromatic camera can
strate that the CC-DFD algorithm combined with also be used for outdoor depth estimation. Figure 15
a chromatic camera can provide accurate depth shows examples of raw depth maps obtained in vari-
estimation and is robust to various scene textures. ous outdoor situations with the CC-DFD algorithm
Note, however, that both bias and standard with 9 × 9 square patches and 50% overlapping. In
deviation degrade after 3.5 m. One can explain this the estimated depth maps, various objects can be
degradation by looking at the theoretical PSF size distinguished, including thin ones like the grid in
variations of the prototype lens presented in Fig. 7. the second example. Note that this kind of repetitive
First, the variation of the PSF size reduces with object misleads the correlation technique, so would
depth, so it is not surprising that the accuracy of be difficult to identify with a stereoscopic device.
depth estimation reduces as well. Besides, the PSF 5. Discussion
of the red channel is below one pixel after 3.5 m,
so this channel no longer gives depth information, A. Robustness of Depth Estimation with Respect to
which reduces the performance. Variation of the Scene Spectrum
D. Depth Maps In the proposed DFD approach, we take advantage of
the interchannel chromatic aberration in order to
1. Comparison with an Active Ranging System for estimate depth from a set of potential calibrated
Indoor Scenes PSFs triplets. However, the object spectrum can have
Figure 14 shows three examples of depth maps an influence on the actual PSF size due to intrachan-
obtained from images acquired with our prototype nel chromatic aberration. One option to reduce the
of chromatic camera using the CC-DFD algorithm PSF variability within a color channel would be to
with 21 × 21 square patches and 50% overlapping. use a sensor with narrowband color filter. However,
As they are indoor scenes, the depth maps obtained this approach would increase the complexity of the
1.5
d,G
d,R
0.5 0.5
a
filter helps to eliminate the outliers in the back-
0 0 ground but does not propagate depth information
2 4 2 4
on homogeneous regions. The second regularization
Depth in m Depth in m
(a) (b) approach consists in optimizing a criterion based on
a Markov random field model of the depth map:
1 X
Ed GLC dp
d,B
a 0.5 p
X
0 ‖yg p − yg q‖2
2 4 λ exp − 1 − δdp ;dq ;
Depth in m p;q∈N p
2σ 2
(c)
(22)
Fig. 17. Weights for the high frequency transfer equation (21) as
a function of depth. where N p is a first-order neighborhood of the pixel p,
dp is the estimated depth value at pixel p, and yg is
the result of color image conversion in grayscale.
clearly on the details of the red channel presented in This energy favors depth jumps located on image
Figs. 16(c)–16(f). A quantitative evaluation of the re- edges. We minimize this criterion using a graphcut
storation gain could be the subject of further studies. algorithm [12,15,39]. Figure 18(b) presents the
C. Depth Map Regularization results obtained with λ 1.1 and σ 4.10−4 . The
parameters have been chosen to propagate informa-
In Fig. 14, we have presented raw depth maps tion over homogeneous regions and lead to a result
obtained with the CC-DFD algorithm. These raw that is close to a depth segmentation.
results present few outliers, and show no depth infor-
mation on homogeneous regions that are excluded 6. Conclusion
from the processing. A regularization approach
can be use to overcome these two issues. Here we In this paper we have proposed a new passive
present preliminary results of two regularization method for depth estimation based on the use of
approaches. The first one, presented in Fig. 18(a), chromatic aberration and a DFD approach. We have
presented an algorithm that estimates depth locally
using a PSFs triplet selection criterion derived from
a maximum likelihood calculation. This algorithm is
based on a modelization of the interchannel scene
correlation, which allows estimation of depth on color
scene patches. Simulated and experimental tests
have illustrated the effectiveness of the proposed al-
gorithm and provide a demonstration of the concept
of chromatic DFD.
There are several perspectives for this work. The
calibration process could be improved so as to reduce
2.5 the number of calibration images required. Regard-
ing the processing, we are currently working on a
parallel implementation of the CC-DFD algorithm.
A more detailed study of image restoration and
2 depth map regularization should also be conducted.
Finally, the design of the chromatic lens could be
optimized to improve the overall performance of
the system using the codesign approach presented
1.5 in [40].
The authors would like to thank F. Bernard and
L. Jacubowiez for fruitful discussions.
1 References
1. A. Pentland, “A new sense for depth of field,” IEEE Trans.
(a) (b) Pattern Anal. Mach. Intell. PAMI-9, 523–531 (1987).
2. Y. Bando, B. Y. Chen, and T. Nishita, “Extracting depth and
Fig. 18. Result of depth map regularization. From up to down:
matte using a color-filtered aperture,” ACM Trans. Graph.
acquired image, raw depth map, regularized depth map: (a) with 27, 1–9 (2008).
a median filter of size 3 × 3, (b) after the minimization of Eq. (22). 3. A. Lumsdaine and T. Georgiev, “The focused plenoptic
The depth labels are in m. Black label corresponds to homogeneous camera,” in Proceedings of IEEE International Conference
regions rejected by the algorithm. on Computational Photography (IEEE, 2009), pp. 1–8.