Structlight

In IEEE Computer Society Conference on Computer Vision and Pattern Recognition
(CVPR 2003), volume 1, pages 195–202, Madison, WI, June 2003

High-Accuracy Stereo Depth Maps Using Structured Light

Daniel Scharstein Richard Szeliski
Middlebury College Microsoft Research
schar@middlebury.edu szeliski@microsoft.com

Abstract
Recent progress in stereo algorithm performance is
quickly outpacing the ability of existing stereo data sets to
discriminate among the best-performing algorithms, moti-
vating the need for more challenging scenes with accurate
ground truth information. This paper describes a method
for acquiring high-complexity stereo image pairs with
pixel-accurate correspondence information using struc-
tured light. Unlike traditional range-sensing approaches,
our method does not require the calibration of the light Figure 1. Experimental setup, showing the digital
sources and yields registered disparity maps between all camera mounted on a translation stage, the video pro-
pairs of cameras and illumination projectors. We present jector, and the complex scene being acquired.
new stereo data sets acquired with our method and demon-
strate their suitability for stereo algorithm evaluation. Our
results are available at http://www.middlebury.edu/stereo/.
solve if noise, aliasing, etc. are not modeled, or too difficult,
e.g., due to complete lack of texture in parts of the scene.
In this paper, we use structured light to uniquely label
each pixel in a set of acquired images, so that correspon-
1. Introduction
dence becomes (mostly) trivial, and dense pixel-accurate
correspondences can be automatically produced to act as
The last few years have seen a resurgence of interest in
ground-truth data. Structured-light techniques rely on pro-
the development of highly accurate stereo correspondence
jecting one or more special light patterns onto a scene,
algorithms. Part of this interest has been spurred by funda-
usually in order to directly acquire a range map of the
mental breakthroughs in matching strategies and optimiza-
scene, typically using a single camera and a single projector
tion algorithms, and part of the interest is due to the exis-
[1, 2, 3, 4, 5, 7, 11, 13, 18, 19, 20, 22, 23]. Random light
tence of image databases that can be used to test and com-
patterns have sometimes been used to provide artificial tex-
pare such algorithms. Unfortunately, as algorithms have
ture to stereo-based range sensing systems [14]. Another
improved, the difficulty of the existing test images has not
approach is to register range data with stereo image pairs,
kept pace. The best-performing algorithms can now cor-
but the range data is usually of lower resolution than the
rectly match most of the pixels in data sets for which correct
images, and the fields of view may not correspond exactly,
(ground truth) disparity information is available [21].
leading to areas of the image for which no range data is
In this paper, we devise a method to automatically
available [16].
acquire high-complexity stereo image pairs with pixel-
accurate correspondence information. Previous approaches
have either relied on hand-labeling a small number of im- 2. Overview of our approach
ages consisting mostly of fronto-parallel planes [17], or set-
ting up scenes with a small number of slanted planes that The goal of our technique is to produce pairs of real-
can be segmented and then matched reliably with para- world images of complex scenes where each pixel is labeled
metric correspondence algorithms [21]. Synthetic images with its correspondence in the other image. These image
have also been suggested for testing stereo algorithm per- pairs can then be used to test the accuracy of stereo algo-
formance [12, 9], but they typically are either too easy to rithms relative to the known ground-truth correspondences.

I-195

Our approach relies on using a pair of cameras and one 8. Optionally crop and downsample the disparity maps
or more light projectors that cast structured light patterns and the views taken under ambient lighting.
onto the scene. Each camera uses the structured light se-
quence to determine a unique code (label) for each pixel. The remainder of this paper is structured as follows.
Finding inter-image correspondence then trivially consists The next section describes the algorithms used to determine
of finding the pixel in the corresponding image that has the unique codes from the structured lighting. Section 4 dis-
same unique code. cusses how view disparities and illumination disparities are
The advantage of our approach, as compared to using a established and merged. Section 5 describes our experimen-
separate range sensor, is that the data sets are automatically tal results, and Section 6 describes our conclusions and fu-
registered. Furthermore, as long as each pixel is illuminated ture work.
by at least one of the projectors, its correspondence in the
other image (or lack of correspondence, which indicates oc- 3. Structured light
clusion) can be unambiguously determined.
In our current experimental setup (Figure 1), we use To uniquely label each pixel, we project a series of struc-
a single digital camera (Canon G1) translating on a lin- tured light images onto the scene, and decode the set of pro-
ear stage, and one or two light projectors illuminating the jected intensities at each pixel to give it a unique label. The
scene from different directions. We acquire images un- simplest kind of pattern to project is a series of single stripe
der both structured lighting and ambient illumination condi- images (light planes) [3, 7, 19], but these require O(n) im-
tions. The ambient illuminated images can be used as inputs ages, where n is the width of the image in pixels.
to the stereo matching algorithms being evaluated. Instead, we have tested two other kinds of structured
Let us now define some terminology. We distinguish be- light: binary Gray-code patterns, and series of sine waves.
tween views – the images taken by the cameras – and illu-
minations – the structured light patterns projected onto the 3.1. Gray codes
scene. We model both processes using planar perspective
projection and use coordinates (x, y) for views and (u, v) Gray-code patterns only contain black and white (on/off)
for illuminations. pixel values, which were the only possibilities available
There are two primary camera views, L (left) and R with the earliest LCD projectors. Using such binary im-
(right), between which correspondences are to be estab- ages requires log2 (n) patterns to distinguish among n lo-
lished. The illumination sources from which light patterns cations. For our projector (Sony VPL-CX10) with 1024×
are projected are identified using numbers {0, 1, . . .}. More 768 pixels, it is sufficient to illuminate the scene with 10
than one illumination direction may be necessary to illumi- vertical and 10 horizontal patterns, which together uniquely
nate all surfaces in complex scenes with occluding objects. encode the (u, v) position at each pixel. Gray codes are
Our processing pipeline consists of the following stages: well suited for such binary position encoding, since only
one bit changes at a time, and thus small mislocalizations
1. Acquire all desired views under all illuminations.
of 0-1 changes cannot result in large code changes [20].
2. Rectify the images to obtain the usual horizontal Decoding the light patterns is conceptually simple, since
epipolar geometry, using either a small number of cor- at each pixel we need only decide whether it is illuminated
responding features [15] or dense 2D correspondences or not. We could for example take two reference images,
(step 4). all-white and all-black, and compare each code pixel with
3. Decode the light patterns to get (u, v) codes at each the average of the two. (With a gray-level projector, we
pixel in each view. could also project a reference image with 0.5 intensity).
Such reference images measure the albedo of each scene
4. Use the unique codes at each pixel to compute corre- point. In practice, however, this does not work well due to
spondences. (If the images are rectified, 1D search can interreflections in the scene and “fogging” inside the pro-
be performed, else 2D search is required.) The results jector (adding a low-frequency average of intensities to the
of this correspondence process are the (usual) view dis- projected pattern), which causes increased brightness near
parities (horizontal motion). bright areas. We have found that the only reliable way of
5. Determine the projection matrices for the illumination thresholding pixels into on/off is to project both the code
sources from the view disparities and the code labels. pattern and its inverse. We can then label each pixel accord-
ing to whether the pattern or its inverse appears brighter.
6. Reproject the code labels into the two-view geometry. This avoids having to estimate the local albedo altogether.
This results in the illumination disparities. The obvious drawback is that twice as many images are re-
7. Combine the disparities from all different sources to quired. Figure 2 shows examples of thresholded Gray-code
get a reliable and accurate final disparity map. images.

I-196

precision for the same number of images). It can also po-
tentially overcome discretization problems that might intro-
duce artifacts at the boundaries of binary patterns [6].
Consider for example projecting a pure white pattern and
a gray-level ramp onto the scene. In the absence of noise
and non-linearities, the ratio of the two values would give us
the position along the ramp of each pixel. However, this ap-
proach has limited effective spatial resolution [11, 22]. Pro-
Figure 2. Examples of thresholded Gray-code im- jecting a more quickly varying pattern such as a sawtooth al-
ages. Uncertain bits are shown in gray. (Full-size leviates this, but introduces a phase ambiguity (points at the
versions of all images in this paper are available at same phase in the periodic pattern cannot be distinguished),
http://www.middlebury.edu/stereo/.) which can be resolved using a series of periodic patterns at
different frequencies [13]. A sine wave pattern avoids the
discontinuities of a sawtooth, but introduces a further two-
way ambiguity in phase, so it is useful to project two or
Unfortunately, even using patterns and their inverses more waves at different phases
may not be enough to reliably distinguish light patterns on Our current algorithm projects sine waves at two differ-
surfaces with widely varying albedos. In our experiments, ent frequencies and 12 different phases. The first frequency
we have found it necessary to use two different exposure has a period equal to the whole (projector) image width or
times (0.5 and 0.1 sec.). At each pixel, we select the ex- height; the second has 10 periods per screen.
posure setting that yields the largest absolute difference be- Given the images of the scene illuminated with these
tween the two illuminations. If this largest difference is still patterns, how do we compute the phase and hence (u, v)
below a threshold (sum of signed differences over all color coordinates at each pixel? Assuming a linear image forma-
bands < 32), the pixel is labeled “unknown” (gray pixels in tion process, we have the following (color) image formation
Figure 2), since its code cannot be reliably determined. This equation
can happen in shadowed areas or for surfaces with very low
albedo, high reflectance, or at oblique angles. Ikl (x, y) = A(x, y)Bkl [sin(2πfk u + φl ) + 1], (1)
The initial code values we obtain by concatenating the
bits from all the thresholded binary images need to be where A(x, y) is the (color) albedo corresponding to scene
cleaned up and potentially interpolated, since the camera pixel (x, y), Bkl is the intensity of the (k, l)th projected pat-
resolution is typically higher than projector resolution. In tern, fk is its frequency, and φl is its phase. A similar equa-
our case, the projector has a 1024 × 768 resolution, and tion can be obtained for horizontal sine wave patterns by
the camera has 2048 × 1536. Since the camera only sees a replacing u with v.
subset of the illuminated scene (i.e., it is zoomed in) and il- Assume for now that we only have a single frequency
lumination pixels can appear larger on slanted surfaces, we fk and let cl = cos φl , sl = sin φl , cu = cos(2πfk u),
get even more discrepancy in resolution. In our setup, each su = sin(2πfk u), and C = A(x, y)B. The above equation
illumination pixel is typically 2–4 camera pixels wide. We can then be re-written (for a given pixel (x, y)) as
clean up the Gray code images by filling small holes caused
by unknown bit values. We then interpolate (integer) code Ikl = C[su cl + cu sl + 1]. (2)
values to get a higher resolution and avoid multiple pixels
with the same code. Interpolation is done in the prominent We can estimate the illuminated albedo value C at each
code direction, i.e., horizontally for u and vertically for v. pixel by projecting a mid-tone grey image onto the scene.
We currently compute a robust average over a sliding 1D The above equation is therefore linear in the unknowns
window of 7 values. The results of the entire decoding pro- (cu , su ), which can be (optimally) recovered using linear
cess are shown in Figure 4a. least squares [10], given a set of images with different
(known) phases φl . (In the least squares fitting, we ignore
3.2. Sine waves any color values that are saturated, i.e., greater than 240.)
An estimate of the u signal can then be recovered using
Binary Gray-code patterns use only two different inten- 1 su
sity levels and require a whole series of images to uniquely u = p−1 (
u tan−1 + m), (3)
2π cu
determine the pixel code. Projecting a continuous function
onto the scene takes advantage of the gray-level resolution where pu = W/fu is the sine period (in pixels) and m is the
available in modern LCD projectors, and can thus poten- (unknown) integral phase wrap count. To solve the phase
tially require fewer images (or alternatively, result in greater wrapping problem, we first estimate the value of u using

I-197

(cu,su)
int line
stra
c on

2πfu

(a): Gray code (b): sine wave
Figure 3. Phase estimation from (cu , su ) least
squares fit. The red dot is the least squares solution
to the constraint lines, and the ellipse around it indi- Figure 4. Computed u coordinates (only low-order
cates the two-dimensional uncertainty. bits are shown).

a single wave (f1 = 1), and then repeat the estimation with 4. Disparity computation
f2 = 10, using the previous result to disambiguate the phase.
Since we are using least squares, we can compute a cer- Given N illumination sources, the decoding stage de-
tainty for the u estimate. The normal equations for the least scribed above yields a set of labels (uij (x, y), vij (x, y)), for
squares system directly give us the information matrix (in- each illumination i ∈ {0, . . . , N −1} and view j ∈ {L, R}.
verse covariance) for the (cu , su ) estimate. We can convert Note that these labels not only uniquely identify each scene
this to a variance in u by projecting along the direction nor- point, but also encode the coordinates of the illumination
mal to the line going through the origin and (cu , su ) (Fig- source. We now describe how high-accuracy disparities can
ure 3). Furthermore, we can use the distance of the fitted be computed from such labels corresponding to one or more
point (cu , su ) from the unit circle as a sanity check on the illumination directions.
quality of our sine wave fit. Computing certainties allows
us to merge estimates from different exposures. At present, 4.1. View disparities
we simply pick the estimate with the higher certainty.
Figure 4b shows the results of recovering the u positions The first step is to establish correspondences between the
using sine patterns. For these experiments, we use all 12 two views L and R by finding matching code values. As-
phases (φ = 0◦ , 30◦ , . . . , 330◦ ) and two different exposures suming rectified views for the moment, this amounts to a
(0.1 and 0.5 sec). In the future, we plan to study how the simple 1D search on corresponding scanlines. While con-
certainty and reliability of these estimates varies as a func- ceptually simple, several practical issues arise:
tion of the number of phases used.
• Some pixels may be partially occluded (visible only in
3.3. Comparison one view).
• Some pixels may have unknown code values in some
Figure 4 shows examples of u coordinates recovered illuminations due to shadows or reflections.
both from Gray code and sine wave patterns. The total num-
• A perfect matching code value may not exist due to
ber of light patterns used is 80 for the Gray codes (10 bit
aliasing or interpolation errors.
patterns and their inverses, both u and v, two exposures),
and 100 for the sine waves (2 frequencies and 12 phases • Several perfect matching code values may exist due to
plus 1 reference image, both u and v, two exposures). Vi- the limited resolution of the illumination source.
sual inspection shows that the Gray codes yield better (less • The correspondences computed from different illumi-
noisy) results. The main reason is that by projecting binary nations may be inconsistent.
patterns and their inverses, we avoid the difficult task of es-
timating the albedo of the scene. Although recovering the The first problem, partial occlusion, is unavoidable and will
phase of sine wave patterns potentially yields higher reso- result in unmatched pixels. The number of unknown code
lution and could be done with fewer images, it is also more values due to shadows in the scene can be reduced by us-
susceptible to non-linearities of the camera and projector ing more than one illumination source, which allows us to
and to interreflections in the scene. establish correspondences at all points illuminated by at
In practice, the time to take the images of all structured least one source, and also enables a consistency check at
light patterns is relatively small compared to that of setting pixels illuminated by more than one source. This is ad-
up the scene and calibrating the cameras. We thus use the vantageous since at this stage our goal is to establish only
Gray code method for the results reported here. high-confidence correspondences. We thus omit all pixels

I-198

whose disparity estimates under different illuminations dis- If the underlying disparities and illumination codes are
agree. As a final consistency check, we establish dispari- correct, this is a fast and stable method for computing M0L .
ties dLR and dRL independently and cross-check for con- In practice, however, a small number of pixels with large
sistency. We now have high-confidence view disparities at disparity errors can strongly affect the least-squares fit. We
points visible in both cameras and illuminated by at least therefore use a robust fit with outlier detection by iterating
one source (see Figures 6b and 7b). the above process. After each iteration, only those pixels
Before moving on, let us consider the case of unrectified with low residual errors are selected as input to the next
views. The above method can still be used, except that a iteration. We found that after 4 iterations with successively
2D search must be used to find corresponding codes. The lower error thresholds we can usually obtain a very good fit.
resulting set of high-quality 2D correspondences can then Given the projection matrix M0L , we can now solve
be used to rectify the original images [15]. Equation (4) for d at each pixel, using again a least-squares
fit to combine the two estimates. This gives us the illumi-
4.2. Illumination disparities nation disparities d0L (x, y) (see Figures 6c and 7c). Note
that these disparities are available for all points illuminated
by source 0, even those that are not visible from the right
The next step in our system is to compute another set
camera. We thus have a new set of disparities, registered
of disparities: those between the cameras and the illumina-
with the first set, which includes half-occluded points. The
tion sources. Since the code values correspond to the im-
above process can be repeated for the other camera to yield
age coordinates of the illumination patterns, each camera-
disparities d0R , as well as for all other illumination sources
illumination pair can be considered an independent source
i = 1 . . . N −1.
of stereo disparities (where the role of one camera is played
by the illumination source). This is of course the idea be-
hind traditional structured lighting systems [3]. 4.3. Combining the disparity estimates
The difference in our case is that we can register these
illumination disparities with our rectified view disparities Our remaining task is to combine the 2N + 2 disparity
dLR without the need to explicitly calibrate the illumina- maps. Note that all disparities are already registered, i.e.,
tion sources (video projectors). Since our final goal is to they describe the horizontal motion between views L and
express all disparities in the rectified two-view geometry, R. The first step is to create combined maps for each of L
we can treat the view disparities as a 3D reconstruction of and R separately using a robust average at pixels with more
the scene (i.e., projective depth), and then solve for the pro- than one disparity estimate. Whenever there is a majority of
jection matrix of each illumination source. values within close range, we use the average of this subset
of values; otherwise, the pixel is labeled unknown. In the
Let us focus on the relationship between the left view L
second step, the left and right (combined) maps are checked
and illumination source 0. Each pixel whose view disparity
for consistency. For unoccluded pixels, this means that
has been established can be considered a (homogeneous)
3D scene point S = [x y d 1]T with projective depth dLR (x, y) = −dRL (x + dLR (x, y), y),
d = dLR (x, y). Since the pixel’s code values (u0L , v0L )
also represent its x and y coordinates in the illumination and vice versa. If the disparities differ slightly, they are ad-
pattern, we can write these coordinates as homogenous 2D justed so that the final set of disparity maps is fully con-
point P = [u0L v0L 1]T . We then have sistent. Note that since we also have disparities in half-
occluded regions, the above equation must be relaxed to
P ∼ M0L S,
= reflect all legal visibility situations. This yields the final,
consistent, and highly accurate pair of disparity maps relat-
where M0L is the unknown 4 × 3 projection matrix of illu- ing the two views L and R (Figures 6d and 7d).
mination source 0 with respect to the left camera. If we let The two final steps are cropping and downsampling. Up
m1 , m2 , m3 denote the three rows of M0L , this yields to this point, we are still dealing with full-size (2048×1536)
images. In our setup, disparities typically range from about
u0L m3 S = m1 S, and 210 to about 450. We can bring the disparities closer to zero
v0L m3 S = m2 S. (4) by cropping to the joint field of view, which in effect stabi-
lizes an imaginary point just behind farthest surface in the
Since M is only defined up to a scale factor, we set m34 = 1. scene. This yields a disparity range of 0–240, and an image
Thus we have two linear equations involving the 11 un- width of 1840. Since most current stereo implementations
known entries of M for each pixel whose disparity and il- work with much smaller image sizes and disparity ranges,
lumination code are known, giving us a heavily overdeter- we downsample the images and disparity maps to quarter
mined linear system of equations, which we solve using size (460 × 384). The disparity maps are downsampled us-
least squares [10]. ing a majority filter, while the ambient images are reduced

I-199

Figure 5. The two image pairs: Cones (left) and Teddy (right).

with a sharp 8-tap filter. Note that for the downsampled im- truth data. Table 1 shows the quantitative performance in
ages, we now have disparities with quarter-pixel accuracy. non-occluded areas (percentage of “bad” pixels with large
A remaining issue is that of holes, i.e., unknown dispar- disparity errors) for two error thresholds t = 1 and t = 2.
ity values, which are marked with a special value. While (We ignore points whose true disparities are unknown.)
small holes can be filled by interpolation during the above The results clearly indicate that our data sets are chal-
process, large holes may remain in areas where no illumina- lenging, yet not unrealistically so. Difficulties posed to
tion codes were available to begin with. There are two main the matching algorithms include a large disparity range,
sources for this: (1) surfaces that are highly specular or have complex surface shapes, textureless areas, narrow occlud-
very low albedo; and (2) areas that are shadowed under all ing objects and ordering-constraint violations. Of the algo-
illuminations. Special care must be taken to avoid both sit- rithms tested, the graph-cut method performs best, although
uations when constructing test scenes whose disparities are it clearly cannot handle some of the complex occlusion sit-
to be estimated with the method described here. uations and some highly-slanted surfaces.

5. Results 6. Conclusion

Using the method described in the previous sections, we In this paper we have developed a new methodology
have acquired two different scenes, Cones and Teddy. Fig- to acquire highly precise and reliable ground truth dis-
ure 5 shows views L and R of each scene taken under ambi- parity measurements accurately aligned with stereo im-
ent lighting. L and R are actually views 3 and 7 out of a total age pairs. Such high-quality data is essential to evalu-
of 9 images we have taken from equally-spaced viewpoints, ate the performance of stereo correspondence algorithms,
which can be used for prediction-error evaluation [21]. which in turn spurs the development of even more accu-
The Cones scene was constructed such that most scene rate algorithms. Our new high-quality disparity maps and
points visible from either view L and R can be illuminated the original input images are available on our web site at
with a single light source from above (see Figure 6a). The http://www.middlebury.edu/stereo/. We plan to add these
exception is the wooden lattice in the upper right quadrant new data sets to those already in use to benchmark the per-
in the image. (Its shadow falls on a planar surface, however, formance of stereo correspondence algorithms [21].
so that the missing disparities could be filled in by interpo- Our novel approach is based on taking stereo image
lation.) For the Teddy we used two illumination directions pairs illuminated with active lighting from one or more
(see Figure 7a). Due to the complex scene, however, several projectors. The structured lighting enables us to uniquely
small areas are shadowed under both illuminations. code each scene pixel, which makes inter-camera corre-
Figures 6 and 7 also show the recovered view disparities spondence much easier and more reliable. Furthermore, the
(b) and illumination disparities (c), as well as the final dis- encoded positions enable the recovery of camera-projector
parity maps combined from all sources (d). The combina- disparities, which can be used as an auxiliary source of in-
tion step not only fills in half-occluded and half-shadowed formation to increase the reliability of correspondences and
regions, but also serves to detect outlier disparities (e.g., due to fill in missing data.
to errors in the projection matrix estimation). We have investigated two different kinds of structured
In order to verify that our stereo data sets are useful for light: binary Gray codes, and continuous sine waves. At
evaluating stereo matching algorithms, we ran several of present, the Gray codes give us more reliable estimates of
the algorithms from the Middlebury Stereo Page [21] on projector coordinates, due mainly to their higher insensitiv-
our new images. Figure 8 shows the results of three al- ity to effects such as photometric nonlinearities and inter-
gorithms (SSD with a 21 × 21 shiftable window, dynamic reflections. In future work, we plan to develop the sine-
programming, and graph cuts) on the cropped and down- wave approach further, to see if we can reduce the to-
sampled image pairs, as well as the corresponding ground- tal number of acquired images necessary to recover high-

I-200

(a) (b) (c) (d)

Figure 6. Left view of Cones (one illumination source is used): (a) scene under illumination (note absence of shadows
except in upper-right corner); (b) view disparities; (c) illumination disparities; (d) ﬁnal (combined) disparity map.
Unknown disparities are shown in black.

(a) (b) (c) (d)

Figure 7. Left view of Teddy (two illumination sources are used): (a) scene under the two illuminations; (b) view
disparities; (c) illumination disparities; (d) ﬁnal (combined) disparity map. Unknown disparities are shown in black.

SSD Dynamic progr. Graph cut Ground truth

Figure 8. Stereo results on cropped and downsampled images: Cones (top) and Teddy (bottom).

I-201

Bad % Cones Teddy [5] C. Chen, Y. Hung, C. Chiang, and J. Wu. Range data acquisi-
t=1 t=2 t=1 t=2 tion using color structured lighting and stereo vision. Image
SSD 17.8% 9.3% 26.5% 12.8% and Vision Computing, 15(6):445–456, 1997.
DP 17.1% 9.8% 30.1% 10.5% [6] Y.-Y. Chuang et al. Environment matting extensions: to-
GC 12.6% 7.0% 29.3% 11.4% wards higher accuracy and real-time capture. In SIGGRAPH
2000, pp. 121–130, 2000.
[7] B. Curless and M. Levoy. Better optical triangulation
Table 1. Performance of SSD, dynamic programming,
through spacetime analysis. In ICCV’95, pp. 987–994, 1995.
and graph cut stereo methods on our data sets. The
[8] J. Davis, R. Ramamoorthi, and S. Rusinkiewicz. Spacetime
table shows the percentage of pixels whose disparity stereo: a unifying framework for depth from triangulation.
error is greater than threshold t for t = 1, 2. In CVPR 2003, 2003.
[9] T. Frohlinghaus and J. M. Buhmann. Regularizing phase-
based stereo. In ICPR’96, vol. A, pp. 451–455, 1996.
[10] G. Golub and C. F. Van Loan. Matrix Computation, third
quality pixel encodings. We also plan to ﬁll in the remaining
edition. The John Hopkins University Press, 1996.
pixels currently marked as unknown in our disparity maps,
[11] G. H¨ usler and D. Ritter. Parallel three-dimensional sensing
a
using a combination of semi-automated and manual meth- by color-coded triangulation. Applied Optics, 32(35):7164–
ods. It may also be possible to co-locate the cameras and 7169, 1993.
projectors using mirrors and to use Zickler et al.’s beautiful [12] W. Hoff and N. Ahuja. Surfaces from stereo: integrating fea-
results on reciprocity to deal with highlights [25]. ture matching, disparity estimation, and contour detection.
While doing our research, we became aware of concur- IEEE Trans. Pat. Anal. Mach. Int., 11(2):121–136, 1989.
rent work aimed at acquiring high-quality correspondences [13] E. Horn and N. Kiryati. Toward optimal structured light pat-
with active illumination that is applicable to both static and terns. In Intl. Conf. Recent Advances in 3D Digital Imaging
dynamic scenes [8, 24]. Instead of decoding the projected and Modeling, pp. 28–35, 1997.
light patterns to yield pixel addresses in the projector, these [14] S. B. Kang, J. Webb, L. Zitnick, and T. Kanade. A multi-
alternative methods simply temporally sum up the corre- baseline stereo system with active illumination and real-time
lation or error measures of all frames to directly compute image acquisition. In ICCV’95, pp. 88–93, 1995.
stereo disparities. This results in a simpler approach that [15] C. Loop and Z. Zhang. Computing rectifying homographies
produces high-quality inter-camera correspondences. Un- for stereo vision. In CVPR’99, volume I, pp. 125–131, 1999.
like our method, however, these techniques are not able to [16] J. Mulligan, V. Isler, and K. Daniilidis. Performance evalu-
ation of stereo for tele-presence. In ICCV 2001, vol. II, pp.
ﬁll in semi-occluded areas using projected disparities.
558–565, 2001.
We close the paper with the following challenge. Can
[17] Y. Nakamura, T. Matsuura, K. Satoh, and Y. Ohta. Occlusion
one devise a comparable (or any) technique to acquire high-
detectable stereo - occlusion patterns in camera matrix. In
quality ground truth data for real-time two-dimensional mo- CVPR’96, pp. 371–378, 1996.
tion? The existence of such data would be of invaluable [18] M. Proesmans, L. Van Gool, and F. Defoort. Reading be-
use to the motion estimation community, just as we hope tween the lines - a method for extracting dynamic 3D with
the data presented here will aid in developing better stereo texture. In ICCV’98, pp. 1081–1086, 1998.
matching algorithms. [19] K. Pulli et al. Acquisition and visualization of colored 3D
objects. In ICPR’98, pp. 11-15, 1998.
Acknowledgments [20] K. Sato and S. Inokuchi. Three-dimensional surface mea-
surement by space encoding range imaging. J. Robotic Sys-
tems, 2:27–39, 1985.
Thanks to Alexander Vandenberg-Rodes who was instru- [21] D. Scharstein and R. Szeliski. A taxonomy and evaluation of
mental in creating the scenes and capturing the images. dense two-frame stereo correspondence algorithms. Intl. J.
Comp. Vis., 47(1):7–42, 2002.
References [22] E. Schubert. Fast 3D object recognition using multiple color
coded illumination. In Proc. IEEE Conf. Acoustics, Speech,
and Signal Processing, pp. 3057–3060, 1997.
[1] G. J. Agin and T. O. Binford. Computer description of curved
[23] P. Vuylsteke and A. Oosterlinck. Range image acquisition
objects. IEEE Trans. Comp., C-25(4):439–449, 1976.
with a single binary-encoded light pattern. IEEE Trans. Pat.
[2] J. Batlle, E. Mouaddib, and J. Salvi. Recent progress in Anal. Mach. Int., 12(2):148–164, 1990.
coded structured light as a technique to solve the correspon-
[24] L. Zhang, B. Curless, and S. M. Seitz. Spacetime stereo:
dence problem: a survey. Pat. Recog., 31(7):963–982, 1998.
shape recovery for dynamic scenes. In CVPR 2003, 2003.
[3] P. Besl. Active optical range imaging sensors. In Jorge L.C.
[25] T. Zickler, P. N. Belhumeur, and D. J. Kriegman. Helmholtz
Sanz, editor, Advances in Machine Vision, pp. 1–63, 1989.
stereopsis: exploiting reciprocity for surface reconstruction.
[4] J.-Y. Bouguet and P. Perona. 3D photography on your desk. In ECCV 2002, v. III, pp. 869–884, 2002.
In ICCV’98, pp. 43–50, 1998.

I-202

Structlight

More Related Content

What's hot (20)

Viewers also liked (8)

Similar to Structlight (20)

Structlight