Projector Displays
Shadow Elimination and Occluder
Light Suppression for MultiProjector Displays
Tat-Jen Cham, James Rehg, Rahul Sukthankar, Gita
Proceedings of Computer Vision and Pattern Recognition,
OTHER INTELLECTUAL PROPERTY RIGHT. Intel products are not intended for use in
medical, life saving, life sustaining applications.
Intel may make changes to specifications and product descriptions at any time, without notice.
Copyright © Intel Corporation 2003
* Other names and brands may be claimed as the property of others.
In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition, Madison, WI, 2003
Shadow Elimination and Occluder Light Suppression for Multi-Projector Displays
Tat-Jen Cham1
James M. Rehg2
Rahul Sukthankar3,4
Gita Sukthankar4
School of Computer Engineering; Nanyang Technological University; Singapore 639798
College of Computing; Georgia Institute of Technology; Atlanta, GA 30332
Intel Research Pittsburgh; 417 S. Craig St.; Pittsburgh, PA 15213
The Robotics Institute; Carnegie Mellon University; Pittsburgh, PA 15213
[email protected], [email protected], [email protected], [email protected]
Two related problems of front projection displays which occur when users obscure a projector are: (i) undesirable shadows cast on the display by the users, and (ii) projected light
falling on and distracting the users. This paper provides a
computational framework for solving these two problems
based on multiple overlapping projectors and cameras. The
overlapping projectors are automatically aligned to display
the same dekeystoned image. The system detects when and
where shadows are cast by occluders and is able to determine the pixels which are occluded in different projectors.
Through a feedback control loop, the contributions of unoccluded pixels from other projectors are boosted in the shadowed regions, thereby eliminating the shadows. In addition, pixels which are being occluded are blanked, thereby
preventing the projected light from falling on a user when
they occlude the display. This can be accomplished even
when the occluders are not visible to the camera. The paper
presents results from a number of experiments demonstrating that the system converges rapidly with low steady-state
1. Introduction
The increasing affordability and portability of high quality projectors has generated a surge of interest in projectorcamera systems. Recent examples include the construction
of seamless multi-projector video walls [1, 2, 3], real-time
range scanning [4] and immersive 3-D virtual environment
generation [5]. In most of these previous systems, cameras
are used to coordinate the aggregation of multiple projectors into a single, large projected display. In constructing a
video wall, for example, the geometric alignment and photometric blending of overlapping projector outputs can be
accomplished by using a camera to measure the keystone
distortions in projected test patterns and then appropriately
pre-warping the projected images. The result is a highly
scalable display system, in contrast to fixed format displays
such as plasma screens.
Our goal is to incorporate projected light displays into
standard physical environments such as classrooms, living
rooms, or kitchens. In these settings, a projector-camera
system could create ubiquitous, interactive displays using
the ordinary visible surfaces in a person’s environment. Dis-
plays could be conveniently located on tabletops, nearby
walls, etc. Users could reposition or resize them using simple hand gestures. Displays could even be “attached” to objects in the environment or be made to follow a user around
as desired. This could be particularly compelling as means
to augment the output capabilities of handheld devices such
as PDAs. In order to realize this vision, two challenging
sensing problems must be solved: (1) Create stable displays in the presence of environmental disturbances such
as changing ambient light and occlusions by the users. (2)
Determine where and when to create displays based on user
In this paper we examine the novel visual sensing challenges that arise in creating stable occlusion-free displays
using projected light in real-world environments. In particular, we address the two problems that arise in frontprojection systems when a user passes between the projectors and the display surface: (1) Shadows cast on the display
surface due to the occlusion of one or more projectors by
the user, (2) Bright light projected on the user, which is often a source of distraction and discomfort. We demonstrate
that these problems can be solved without accurate 3-D localization of projectors, cameras, or occluders, and without
accurate photometric calibration of the display surface. The
key is a display-centric camera feedback loop that rejects
disturbances and unmodeled effects.
Our system uses multiple, conventional projectors which
are positioned so that their projections overlap on the selected display surface. 1 The resulting system produces
shadow-free displays even in the presence of multiple, moving occluders. Furthermore, projector light cast on the occluders is suppressed without affecting the quality of the
display as shown in figure 5.
2. System Overview
Our system comprises a number of projectors which are
aimed at a screen such that their projection regions overlap
and a camera which is positioned such that can view the entire screen. During normal functioning, the system displays
a high quality, dekeystoned image on the screen. When
users walk between the projectors and the screen, shadows
1 In the event that not enough fixed projectors are available for a given
environment, a projector-mirror system such as [6] could be used.
are cast on the screen. These shadows can be classified as
umbral when all projectors are simultaneously occluded, or
penumbral when at least one projector remains unoccluded.
The system eliminates all penumbral shadows cast on the
screen,2 as well as suppressing projector light falling on the
occluders. This enables the system to continue presenting
a high quality image without projecting distracting light on
users. See figure 1 for the setup.
Shadow 2
Shadow 1
in each projector to be pre-warped so as to create a desired
projected display that is aligned with the screen. Additionally, the images should be warped quickly and efficiently.
The photometric issues are the accurate and fast computation of the desired pixel intensities in each projector so as to
eliminate shadows and suppress occluder illumination. This
involves occlusion detection based on camera input and correctly adapting the projector output to achieve the necessary
goals. These two classes of problems are addressed in sections 3 and 4 respectively.
Display surface (S)
3. Multi-Projector Alignment
Camera (C)
Projector (P1)
Projector (P2)
Figure 1: An overhead view of the multi-projector display system.
Several projectors (P1 , P2 ) are placed such that their projection areas converge onto the display surface (S). A camera (C) is positioned so that S is clearly visible in its field of view. Once the
homographies relating Pi , C, S are automatically determined, the
projectors combine their pre-warped outputs to create a single highquality image on S. The system simultaneously removes shadows
on the display and light projected on the occluders, even when the
occluders are not visible in the camera.
A classical approach to designing such a system would
be to localize the projectors, camera, display, and occluders
in 3-D. This would make it possible to predict the shadowed
region and the projector pixels which are being occluded.
Given further information about the reflectance properties
of the surface, the optimal compensation could be determined and applied. This approach has a strong disadvantage that any errors in localization and sensing have a direct
impact on the quality of the display. Particularly in the case
of tracking occluders, significant errors are likely. This paper explores the alternative approach of defining the desired
screen output and using geometric and photometric compensation, inside a tight visual feedback loop, to correct for
errors and disturbances.
The two classes of problems to be addressed in our system are: (i) geometric, and (ii) photometric. The geometric problems relate to computation of the spatial correspondences between pixels in the projectors and the projected
display on the screen. The projectors should be accurately
and automatically calibrated to the screen, to the camera
and to each other. The calibration should enable the images
2 By
definition, pixels in an umbral shadow are blocked from every projector and cannot be removed. Umbral shadows can be minimized by increasing the number of projectors and by mounting the projectors at highlyoblique angles.
As shown in figure 1, several projectors are placed so that
their projection areas all converge onto a display surface
S. The goal is to combine the light from the projectors to
create a single, sharp image on S. Clearly, one cannot simply project the same raw image simultaneously through the
different projectors; not only does a given point on S correspond to very different pixel locations in each projector, but
the image produced on S from any single projector will be
distorted (since the projectors are off-center to S).
The geometric relation between the projector output, the
display on the screen and the camera input can be modeled
via 2-D planar homographies. The homography for each
camera-projector pair T c,Pi can be determined by projecting
a rectangle from the given projector into the environment.
The coordinates of the rectangle’s corners in projector coordinates (xi , yi ) are known a priori, and the coordinates of
the corners in the camera frame (X i , Yi ) are located using
standard image processing techniques. 3
The display area is either automatically determined using the camera, or interactively specified by the user. The
former case requires the display surface to be a white, rectangular projection screen against a contrasting background.
Such a screen shows up clearly in the camera image as a
bright quadrilateral, and can be unambiguously identified
by the automatic calibration process.
Alternatively, the user may interactively specify the display area by manipulating the outline of a projected quadrilateral in any projector until it appears as a rectangle of
the desired size on the display surface. This directly specifies the homography between the selected projector and the
screen TPi ,s ; the outline of the selected rectangle is then
be detected in the camera image to determine the camera
to screen homography T c,s and to automatically align all of
the other projectors.
More recently, we presented a system [8] that is capable
of detecting physical markers and update homographies on
the fly. The system is also capable of maintaining a stable
display even when projectors and cameras are shifted.
3 Hough-transform line-fitting [7] locates the edges of the quadrilateral,
and its corner coordinates are given by intersecting these lines.
The projector-screen homographies T Pi ,s model the geometric distortion (keystone warping) that is induced when
an image is projected from an off-center projector P i . This
distortion can be corrected by projecting a pre-warped image, generated by applying the inverse transform T P−1
i ,s
the original image. 4 Since TP−1
Pi ,s
i ,s
that the pre-warping also aligns the images from different
projectors so that all are precisely projected onto S. Applying the homographies derived from camera images, a
multi-projector array can thus be efficiently configured to
eliminate keystoning distortions and double images on the
display surface.
4. Photometric Framework for MultiProjector Display
After the projectors have been geometrically aligned, we
can easily determine which source pixels from the projectors contribute to the intensity of an arbitrary screen pixel.
In the following analysis, we assume that the contributions
are at some level additive. Given N projectors, the observed
intensity Zt of a particular screen pixel at time t may be expressed by
Zt = C k1t S1 (I1t ) + · · · + kN t SN (IN t ) + A ,
where Ijt is the corresponding source pixel intensity set in
projector j at time t, S j (·) is the projector to screen intensity
transfer function, A is the ambient light contribution which
is assumed to be time invariant, C(·) is the screen to camera
intensity transfer function and k jt is the visibility ratio of
the source pixel in projector j at time t. Note that all the
variables and functions also depend on the spatial position
of the screen pixel, but this is omitted from the notation
since we will consider each pixel in isolation. See figure 2.
When occluders obstruct the paths of the light rays from
some of the projectors to the screen, Z t diminishes and
shadows occur. This situation is quantitatively modeled via
the visibility ratios, which represent the proportion of light
rays from corresponding source pixels in the projectors that
remain unobstructed. If the projectors were modeled as
point-light sources, occluders would block either none or
all of the light falling on a given pixel from any particular
projector; therefore, k jt would be a binary variable. However, this assumption is not valid in real-world conditions.
Our system must cope with partial occluders (created by objects near the projector) that cast fuzzy-edged shadows on
the screen. In these cases k jt denotes the degree of occlusion of projector j for the given pixel.
4 In our current system, this pre-warp is efficiently implemented using
the texture-mapping operations available in standard 3-D graphics hardware.
full occluder
Projector 3
Partial occluder
Projector 1
Projector 2
Figure 2: Photometric framework. This diagram illustrates equation (1), in which the observed display intensity Zt is related to the
combination of projector source pixels Ijt and the corresponding
visibility ratios kjt . The visibility ratios vary accordingly with nonocclusion, partial and full occlusion.
4.1. Occlusion detection
Rather than locating occluders by tracking objects in the
environment, the system focuses exclusively on detecting
deviation of the observed intensities on the screen from the
desired intensities when occluders are not present. The major cause of deviation is occlusion, although deviation can
also occur because of changes in ambient lighting, projector failure, etc. Our system can handle all of these problems
(as discussed in the next section). No assumptions are made
about the locations, sizes or shapes of occluders.
Mathematically, the desired intensity of a particular
screen pixel may be represented by Z 0 . This may be obtained in the initialization phase when the system projects
each presentation slide and captures several camera images
of the projected display while occluders are absent. As
an occluder is introduced in front of projector k to create penumbral shadows, the visibility ratio k jt decreases,
such that kjt < 1. Hence Zt < Z0 . These deviations in
the screen can be detected via a pixel-wise image difference between current and reference camera images to locate
shadow artifacts.
4.2. Iterative Photometric Compensation
Our system handles occluders by
1. compensating for shadows on the screen by boosting
the intensities of unoccluded source pixels; and
2. removing projector light falling on the occluder by
blanking the intensities of occluded source pixels.
The degrees-of-freedom available to us are the source pixels
Ijt which may be changed. Hence for a shadowed screen
pixel where Zt < Z0 , we ideally want to compensate for the
shadow (i.e. setting Z t+1 = Z0 ) by (i) increasing I j(t+1) to
be larger than I jt if kjt = 1, and (ii) reducing I j(t+1) to
zero if kjt < 1.
However, it is very difficult to accurately model C(·) and
Sj (·). Even if we know the exact values for the ambient
lighting and visibility ratios, it is almost impossible to update the source pixels such that in one time step the shadows
are eliminated. Fortunately, we expect C(·) and S j (·) to be
positive monotonic, and an iterative negative feedback loop
can be used to compute I 1t , . . . , IN t required to minimize
Zt − Z0 .
The advantages of such a system are:
• it does not require explicit modeling of C(·) and S j (·),
• it does not require explicit measurement of the visibility ratios kjt ,
• it is able to handle slowly varying ambient light.
As in [9], the change in the intensity of each source pixel
in each projector is controlled by the alpha value associated
with the pixel:
Ijt = αjt I0 ,
where I0 is the original value of the source pixel (i.e. pixel
value in the presentation slide) and is the same across all
projectors, while α jt , 0 < αjt < 1 is the time-varying,
projector-dependent alpha value. The alpha values for the
source pixels in one projector is collectively termed the alpha mask for the projector.
The earlier system described in [9] can compensate for
shadows but is incapable of suppressing projected light
falling on the occluder. In particular, that simpler method
cannot distinguish between the contributions of individual
projectors. Instead, all projectors boost their pixel intensities for each occluded region. This has two undesirable
consequences: (1) bright “haloes” may appear around eliminated shadows, particularly when occluders are in motion;
and (2) the amount of distracting light projected on users
is increased rather than reduced by the system. This motivates the need for a more complex solution where the alpha
masks are different for different projectors.
The approach adopted here is to design components
which separately handle the problems of shadow elimination and occluder light suppression, and integrate them into
a complete system. These are discussed in the following
4.3. Shadow Elimination
Eliminating shadows involve increasing values for corresponding source pixels. The shadow elimination (SE) component of the system is based on
(∆αjt )SE = −γ(Zt − Z0 ),
where ∆αjt = αj(t+1) − αjt is change of αjt in the next
time-frame, and γ is a proportional constant. This component is a simple, linear feedback system.
4.4. Occluder Light Suppression
Suppressing projector light falling on occluders involve diminishing the source pixels corresponding to the occluded
light rays. We determine whether a source pixel is occluded
by determining if any changes in the source pixel result in
changes in the screen pixel. However, since there are N
possible changes of source pixel intensities from N projectors but only one observable screen intensity, we need to
probe by varying the source pixels in different projectors
separately. This cyclical probing results in a serial variation
of the projector intensities.
The light suppression (LS) component of the system is
based on
∆α2j(t−N )
(∆αjt )LS = −β
∆Zt2 + where ∆Zt = Zt − Zt−N is the change in the screen pixel
intensity caused by the change of alpha value ∆α j(t−N ) in
the previous time frame when projector j is active, and β
is a small proportional constant and is a small positive
constant to prevent a null denominator.
The rationale for (4) is that if the change in α jt results in
a corresponding-sized change in Z t , the subsequent change
in αjt will be relatively minor (based on a small β). However if a change in α jt does not result in a change in Z t , this
implies that the source pixel is occluded. The denominator
of (4) approaches zero and α jt is strongly reduced in the
next time frame. Hence occluded source pixels are forced
to black.
Note that the probe technique must be employed during
shadow elimination as well. In particular, the system must
be able to discover when a pixel which was turned off due
to the presence of an occluder is available again, due to the
occluder’s dissapearance. This constraint is smoothly incorporated into our algorithm.
4.5. Integrated System for Shadow Elimination and Occluder Light Suppression
The integrated iterative feedback system combines (3) and
(4) to get
∆αjt = (∆αjt )SE + (∆αjt )LS .
The alpha values are updated within limits such that
if αjt + ∆αjt > 1,
if αjt + ∆αjt < 0,
αjt =
αjt + ∆αjt , otherwise.
The following synthetic example illustrates the system. For
a particular screen pixel at a typical steady state when
shadows have been eliminated, suppose the corresponding
source pixel at projector 1 has α 1t = 0 and the source
pixel at projector 2 has α 2t = 1. At this state, Zt = Z0
and ∆αjt = 0 via (5). If source pixel 2 is suddenly occluded and source pixel 1 is unoccluded, then Z t < Z0
because source pixel 1 is still black. However, ∆α 1t becomes dominated by (∆α 1t )SE which forces source pixel
1 to be bright. On the other hand, ∆α 2t becomes dominated
by (∆α2t )LS since the screen pixel does not change when
αjt is changed. This forces source pixel 2 to be dark. Note
that even when source pixel 2 becomes unoccluded, nothing changes if source pixel 1 remains unoccluded since the
shadows have already been satisfactorily eliminated. See
figure 3. This particularly illustrates the hysteresis effect in
which source pixels are not boosted or blanked until new
shadows are created – the system does not automatically
return to an original state, nor change as a result of deocclusion.
Figure 4 illustrates the algorithm. During its initialization phase (when the scene is occluder-free) the system
projects each presentation slide and captures several camera
images of the projected display. These images are pixelwise averaged to create a reference image for that slide,
and this image represents the desired state of the display
(figure 4, reference image). The goal of occlusion detection is to identify regions in the current image that deviate
from this ideal state. During operation, the system camera continuously acquires images of the projected display
which may contain uncorrected shadows. The comparison
between the observed images and the reference image facilitates the computation of the alpha masks for individual
projectors through (5). These are merged with the presentation slide in the screen frame of reference, followed by further warping into the projector frame of reference. These
projected images from all projectors optically blend to form
the actual screen display.
Computation for Projector 1
Source pixel intensity from projector 1
Source pixel intensity from projector 2
Alpha mask (in camera frame)
Projector-warped slide
Alpha-combined slide
Alpha value
Screenprojector 1
Projector 1
Reference image
both projectors
projector 1
both projectors
projector 2
Raw slide
Camera images
Figure 3: Synthetic example of transitions in projector source pixel
intensities. This graph shows the intensity transition of two corresponding projector source pixels over time, subject to four events of
occlusions and deocclusions. Note the hysteresis effect in which the
source pixels are not boosted or blanked until new occlusion events
Screenprojector 2
Projector 2
Alpha mask (in camera frame)
Alpha-combined slide
Projector-warped slide
Computation for Projector 2
Multiple projected images optically blend into a single, shadow-free display
with occluder light suppression
4.6. System Details
Most of the computation is carried out in the camera frame
of reference. The differences between the observed and desired intensities Zt − Z0 for all screen pixels are obtained
simply by subtracting the current camera image from the
reference camera image. Similarly, the alpha values of the
source pixels within the same projector is collectively the
alpha mask for the projector. The alpha masks are also computed in the camera frame of reference and must be transformed into the screen frame of reference before they can be
applied; this is done using the camera-screen homography
Tc,s discussed in section 3.
Applying the alpha masks to the current slide is straightforward. For each projector, the corresponding alpha mask
is warped and replaces the alpha channel of the slide image. The slide is then pre-warped for the projector (using its
particular screen-to-projector homography) and displayed.
This is done separately for every projector.
Figure 4: System summary diagram. This diagram summarizes the
shadow elimination and occluder light suppression system. See text
for details.
Since we do not have good photometric models of the
environment, projectors or camera, we cannot predict precisely how much light is needed to remove a shadow. However, the iterative feedback loop used to update the alpha
mask allows us to avoid this problem: the system will continue adding light to shadowed regions until the region appears as it did in the reference image. Similarly, the system
will blank projector source pixels which are occluded and
do not affect the observed images. This approach has additional benefits. For instance, the system does not require an
accurate photometric model of the shadow formation process to correct for occlusions with non-binary visibility ratios, e.g., the diffuse shadow created by a hand moving near
the projector. The drawback to such an iterative technique
is that the alpha mask can require several iterations to con-
verge; in practice, shadows are eliminated in approximately
5–7 iterations.
To reduce the effects of camera noise and minor calibration errors, we apply a 5×5 spatial median filter to the
difference image. A negative value in a difference image
pixel means that the corresponding patch on the screen was
under-illuminated in the current image.
5. Results
The experiments described in this section are based on
the following implementation. Images are acquired using an NTSC camera (640×480) attached to a PCI digitizer; the output subsystem consists of two Compaq MP2800 XGA-resolution (1024×768) DLP microportable projectors driven by a dual-headed graphics card; the software
runs on a standard workstation, and image warping exploits
OpenGL texture-mapping primitives. The projectors are positioned on either side of a whiteboard and the camera is
mounted at an extreme angle to ensure occlusion-free coverage of the display surface The location, orientation and optical parameters of the camera and projectors are unknown.
Note that the cost of shadow elimination is the use of
redundant projectors. This means that at any point in time
there are pixels on one or more projectors that are not being
utilized because they fall outside the display surface or are
occluded. We feel this is a small price to pay, particularly in
comparison to the large costs, in either expense and required
space, for other display technologies such as rear projection
or plasma. Fortunately, portable projectors are becoming
increasingly affordable as their image quality improves and
their weight decreases.
Some images of the system in action are shown in figure 5. Note the accuracy of the multi-projector alignment.
In multi-projector display blending, any inaccuracies in the
homography estimation lead to misalignment in projected
images creating glaring artifacts such as double-images. In
figure 5, it can be seen that the multi-projector display is as
sharp as the single-projector display. Additional quantitative measurements are discussed in the following sections.
5.1. Steady State Shadow Elimination
The qualitative results of shadow elimination as shown in
figure 5 are confirmed by experiments comparing our system’s performance with passive one- and two-projector systems. A series of slides were displayed using the three systems, first without occluders and then with static occluders in the scene. Images of the display region were captured using a tripod-mounted camera. We computed a sumsquared-difference (SSD) of the grey-level intensities (over
the pixels corresponding to the display region in the camera image) for each slide pair (with and without occluders).
Figure 5: Comparison between different projection systems. These
images were taken from an audience member’s viewpoint: (a) Two
aligned projectors, passive; (b) Two aligned projectors with shadow
elimination only; (c,d) Two aligned projectors with shadow elimination and occluder light suppression. Note that the user’s face is
harshly illuminated in (a) and (b), while completely dark in (c) and
Table 1 summarizes these results. As expected, the (umbra) shadows in the single-projector display are the major
source of error. We see a significant improvement with two
projectors since the majority of occlusions become penumbral. The best results are achieved when these penumbral
occlusions are actively eliminated using our system.
One projector
Two projectors
Shadow elimination
Sh. elim. + occ. light supp.
(i) SSD error per
pixel over display
(norm. intensities)
(ii) Avg. normalized
intensity measured
on occluder’s face
148.0 × 10−3
12.8 × 10−3
0.242 × 10−3
2.5 × 10−3
Table 1: The result shows (i) average SSD error between an unoccluded reference image and the same scene with a person occluding both projectors, and (ii) normalized intensity of light on the
occluder, averaged over the visible pixels. A two-projector display
improves the display by simply muting the shadows. Our system
not only removes the shadows entirely as shown in figure 5, but
also eliminates the distracting illumination on the user’s face.
Average Pixel Intensity on Occluder
5.2. Steady State Occluder Light Suppression
We compare the amount of projector light falling on the occluder by measuring the average pixel intensity of the user’s
face in images taken from an audience viewpoint. The passive single-projector system brightly illuminates the user’s
face.5 In the passive two-projector system, each projector
operates at a reduced brightness; the measured average intensity is lower because the user’s face is illuminated from
two directions and some of the light from the second projector illuminates areas of the face that are hidden from the
audience camera. The shadow-elimination system from [9]
also uses two projectors, but each projector increases illumination for pixels in the shadowed region, creating a very
bright patch of light on the occluder. The system described
in this paper dramatically suppresses occluder illumination
as can be seen from figure 5 and Table 1.
5.3. System Response
Since our algorithm uses an iterative technique, we wanted
to confirm that the method converges quickly to a stable
value without excessive oscillation or overshoot. We evaluate the system response to occluders by measuring the timevariation of (i) the average SSD per pixel between the observed and reference screen images, and (ii) the average
pixel intensity on the occluder’s face. The response is measured by suddenly introducing an occluder into the scene
and recording the observed state of the display from that instant. The measurements were taken with a video camera
placed in the audience area.
Figures 6 (a) and (b) present system traces for both
experiments. In each graph, one trace shows the system
from [9] while the other plots the response of the combined
shadow elimination and occluder light suppression system.
The combined system takes much longer to converge for
two reasons. First, the system is more complex and runs
more slowly. Each iteration of the algorithm takes approximately 3–4 times longer than a comparable iteration for the
former system; this is easily seen from the length of the
steps in the graphs. Second, the combined system requires
cyclical projector probing to determine the occluded projector. Consequently, an N projector system requires N times
longer to converge.
The initial SSD error for the two systems differ because
the former starts with both projectors on (equivalent to a 2projector passive system) while the combined system starts
with one projector on, and the other projector off. Thus,
when the occluder is initially introduced, the shadows for
the combined system are as dark as umbral occlusions. The
5 This is identical to a conventional front-projection setup. Many users
find this blinding light very irritating and prefer rear-projection systems for
this reason.
Average SSD per pixel
Shadow Elimination +
Occluder Light Suppression
Shadow Elimination Only
Shadow Elimination + Occluder Light Suppression
Shadow Elimination Only
Time frames
(a) Shadow-Elimination Response
Time frames
(b) Occluder Light Response
Figure 6: System responses for comparison of previous[9] and current systems. (a) Responses for shadow elimination, based on average SSD error per pixel over time. (b) Responses for occluder
illumination, based on average pixel intensity of region containing
occluder’s face. Although the current system is slower, there is
substantial occluder light suppression and the steady-state shadow
elimination error is comparable.
SSD errors for both systems drop quickly with each iteration and the final errors are comparable.
Figure 6(b) demonstrates the clear benefits of the combined system. The introduction of the occluder causes an
increase in brightness on the user’s face for the former system while the combined system quickly removes the undesirable light.
Another advantage of our current system is that it is
significantly more resistant to a moving occluder. In [9],
the slight swaying motion of a standing occluder creates
very noticeable “halo” effects due to the occluder’s leading edge creating a shadow while the trailing edge creating
an inverse-shadow (when occluded source pixels in a projector are unoccluded). In our system, the occluded source
pixels are blanked; hence no inverse-shadows are created by
deocclusion. Additionally, the hysteresis effect ensures that
source pixel states are not changed until new shadows are
created. This means that source pixels cyclically occluded
by the span of the occluder swaying motion eventually become blanked and no new shadows (or artifacts) are caused.
A benefit of using a feedback system (as opposed to
a photometric model approach) is that the system is surprisingly robust. For instance, if one of the projectors in
the multi-projector array were to fail, the remaining projectors would automatically brighten their images to compensate. Furthermore, the overall brightness of the entire
multi-projector array can be changed simply by adjusting
the camera aperture.
5.4. Predictive Photometric Correction with a
Two Projector System
Our system solves two problems through feedback control
of projected light: (i) iterative feedback avoids the need for
accurate photometric models, and (ii) cyclical probing circumvents the need to instantly determine which projectors
are occluded for each pixel.
The disadvantage of this approach is the rate of conver-
gence which is approaching but not yet reached a sufficient
speed for normal use.
In order to study a special case where these two problems
are solved through simpler means, we have implemented a
two projector binary switching system, where the projector
contributive output on the screen has been equalized manually. This resolves the issue of photometric models. Additionally, each pixel on the screen is illuminated by only one
projector at any one time. Hence if there is any observed
deviation in pixel value from the reference pixel value, it
may be deduced that the illuminating projector is occluded
without active probing. The system immediately blanks the
pixel contribution from the current illuminating projector,
and turns on the contribution to the other projector. This
results in a binary switching process that is extremely fast.
The exact solution is obtained in a single time frame, and
in our experiments the current implementation eliminates
shadows and suppresses occluder light at about 3 Hz.
light on the occluders. This system enables immersive, omnipresent displays to be generated in any general indoor location, which is impractical or impossible with monitors,
plasma displays and rear-projection systems.
In the future, we plan to extend the system in several
ways. In addition to increasing the frame rate at which the
system operates, we will incorporate multiple cameras into
the visual feedback loop. This will increase the system’s reliability in cases where a single camera’s view of the screen
is also partially occluded. We aim to further develop the
connections between our system and the visual servoing literature [13, 14]. Our current feedback loop can be viewed
as a proportional controller and we hope to explore extensions to PID control. We are also developing user-interface
techniques for controlling and adjusting virtual displays using hand gestures. In particular, we are exploring shadow
detection as a means to support touch-based interaction with
the virtual display.
6. Related Work
Research in the area of camera-assisted multi-projector displays is becoming more popular, particularly in the context
of seamless video walls [1, 2, 3, 5]. However, there has
been little work in using redundant projector configurations
to create stable forward-projection displays. In [9], a simple
camera feedback system is used to generate a single alpha
mask that is appropriately warped and applied to every projector in the display to eliminate shadows. Although this
technique works in real-time, it generates significant additional light on the presenter. Jaynes et al. [10] removed
shadows without a feedback loop, but the resulting algorithm is very slow. We believe that our results in removing
the projected light falling on occluding objects are unique.
A simple camera feedback system, related to the one presented here, was used by [3] to adjust projector illumination
for uniform blending in the overlap region of a video wall.
In [6] a projector-mirror system is used to steer the output of
a single projector to arbitrary locations in an environment.
The Shader Lamps system [11] uses multiple projectors and
a known 3-D model to synthesize interesting visual effects
on 3-D objects. The self-calibration techniques used in this
paper were adopted from [12], where they were applied to
the task of automatic keystone correction for single projector systems.
7. Conclusions and Future Work
This paper describes a practical system for creating a stable, occlusion-free display on ordinary wall surfaces using
multiple overlapping front projectors and a camera. The
system automatically aligns the projected image to the display, compensates for shadows cast on the display due to
projector occlusion, and also removes unwanted projector
[1] M. Hereld, I. Judson, and R. Stevens, “Introduction to building
projection-based tiled displays,” Computer Graphics and Applications, vol. 20, no. 4, 2000.
[2] K. Li, H. Chen, Y. Chen, D. Clark, P. Cook, S. Daminakis,
G. Essl, A. Finkelstein, T. Funkhouser, A. Klein, Z. Liu, E. Praun,
R. Samanta, B. Shedd, J. Singh, G. Tzanetakis, and J. Zheng, “Building and using a scalable display wall system,” Computer Graphics
and Applications, vol. 20, no. 4, 2000.
[3] R. Surati, A Scalable Self-Calibrating Technology for Seamless
Large-Scale Displays. PhD thesis, EECS, MIT, 1999.
[4] O. Hall-Holt and S. Rusinkiewicz, “Stripe boundary codes for realtime structured light range scanning of moving objects,” in Proc.
ICCV, 2001.
[5] R. Raskar, M. Brown, R. Yang, W. Chen, G. Welch, H. Towles,
B. Seales, and H. Fuchs, “Multi-projector displays using camerabased registration,” in Proc. IEEE Visualization Conf., 1999.
[6] C. Pinhanez, “The Everywhere display,” in Proc. Ubiquitous Computing Conf., 2001.
[7] D. Ballard and C. Brown, Computer Vision. Prentice-Hall, 1982.
[8] J. Rehg, M. Flagg, T. Cham, R. Sukthankar, and G. Sukthankar, “Projected light displays using visual feedback,” in Proc. ICARCV, 2002.
[9] R. Sukthankar, T.-J. Cham, and G. Sukthankar, “Dynamic shadow
elimination for multi-projector displays,” in Proc. CVPR, 2001.
[10] C. Jaynes, S. Webb, R. Steele, M. Brown, and W. Seales, “Dynamic
shadow removal from front projection displays,” in Proc. IEEE Visualization, 2001.
[11] R. Raskar, G. Welch, and K.-L. Low, “Shader Lamps: Animating
real objects with image-based illumination,” in Proc. Eurographics
Workshop on Rendering, 2001.
[12] R. Sukthankar, R. Stockton, and M. Mullin, “Smarter presentations: Exploiting homography in camera-projector systems,” in Proc.
ICCV, 2001.
[13] P. Corke, “Visual control of robot manipulators — a review,” in Visual Servoing (K. Hashimoto, ed.), World Scientific, 1994.
[14] T. Uhlin, P. Hordlund, A. Maki, and J. Eklundh, “Towards an active
visual observer,” in Proc. ICCV, 1995.
Was this manual useful for you? yes no
Thank you for your participation!

* Your assessment is very important for improving the work of artificial intelligence, which forms the content of this project

Download PDF