Multi-Projector Displays Using Camera-Based

Multi-Projector Displays Using Camera-Based
Multi-Projector Displays Using Camera-Based Registration
Ramesh Raskar∗, Michael S. Brown†, Ruigang Yang, Wei-Chao Chen,
Greg Welch, Herman Towles, Brent Seales† , Henry Fuchs
Department of Computer Science
University of North Carolina at Chapel Hill
Conventional projector-based display systems are typically
designed around precise and regular configurations of projectors and display surfaces. While this results in rendering
simplicity and speed, it also means painstaking construction
and ongoing maintenance. In previously published work, we
introduced a vision of projector-based displays constructed
from a collection of casually-arranged projectors and display
In this paper, we present flexible yet practical methods
for realizing this vision, enabling low-cost mega-pixel display systems with large physical dimensions, higher resolution, or both. The techniques afford new opportunities to
build personal 3D visualization systems in offices, conference
rooms, theaters, or even your living room. As a demonstration of the simplicity and effectiveness of the methods that
we continue to perfect, we show in the included video that a
10-year old child can construct and calibrate a two-camera,
two-projector, head-tracked display system, all in about 15
CR Categories: I.3.3 [Computer Graphics]: Picture/Image
Generation -Digitizing and scanning, Display algorithms, Viewing algorithms; I.3.7 [Computer Graphics]: Three-Dimensional
Graphics and Realism - Virtual reality; I.4.1 [Image Processing
and Computer Vision]: Digitization and Image Capture - Imaging
geometry, Camera calibration, Sampling, Scanning; I.4.8 [Image
Processing and Computer Vision]: Scene Analysis - Range data,
Surface fitting, Tracking; B.4.2 [Input/Output and Data Communications] Input/Output Devices - Image display.
Additional Keywords: display, projection, spatially immersive
display, panoramic image display, virtual environments, intensity
blending, image-based modeling, depth, calibration, auto-calibration,
structured light, camera-based registration.
∗ {raskar,
ryang, ciao, welch, herman, fuchs}
† {mbrown,seales}
The lure of building a single logical display from a set of
individual light projectors is inspired by the promise of very
high-resolution displays with large areas of coverage together
with affordable components. Such large field of view displays are traditionally created using a well-configured set
of projectors so that they do not create keystoning and are
physically aligned to match the neighboring projectors. Such
panoramic displays for flight simulations, virtual reality and
visualization create spectacular imagery [1, 2, 3, 4, 5, 6].
However, the physical construction of these idealized systems requires considerable space and constant attention and
adjustment by trained personnel. The large amount of computing/rendering power that is now available and ever increasing allows us to consider a different set of tradeoffs.
With additional computation costs, we can design a generalized solution - one that accommodates display in normal
rooms using a casual placement of projectors.
Figure 1: (a) Our display techniques allow for an arbitrary
configuration of projectors and displays. (b) Example of
arbitrary projector overlaps before calibration. (c) Viewer
in the final display environment. (see color plate)
One central problem to be solved in achieving seamless
imagery with multiple projectors is that of geometric registration. Broadly, geometric registration is the alignment of
image features on the target display surface across individual
light-projector boundaries. Correct registration, together
with correct image intensity blending of the overlapped image areas, makes these boundaries visually insignificant, and
creates a powerful immersive effect for the viewer. The effect is broken when the panorama contains seams or obvious
gaps and overlaps.
Few panoramic display systems have explored the issue of
maintaining compelling geometric registration between overlapping projected images for a moving user. We achieve
geometric registration by recovering a 3D representation of
the display environment. This includes taking into account
three critical components:
• the configuration of the set of projectors,
• the geometry of the display surface,
• and the location of the viewer
Figure 2: Examples of panoramic image display environments
A common way to represent the many components of the
panoramic display problem is to consider a world coordinate
system (WCS) defined in the viewer’s physical space to encompass viewer position and a representation of the screen
surface. Representations of each of the components of the
system are expressed within this WCS: the 3D model to be
rendered, models of the light projectors, models of the cameras, and a model of the display surface. At the minimum
scale, the problem is to consider a single camera and projector. Scaling the environment means that the number of projectors and cameras are increased in order to obtain higher
resolution and larger areas of physical screen coverage. Any
method to create a unified display from camera/projector
components must be based on the geometric relationships
between component models.
The full-scale panoramic image display problem is to compute images for the frame buffers of each projector such that
the images combine properly on the display surface to produce geometrically correct, seamless imagery for the viewer.
When we assume an idealized and completely known geometry for display configuration (all components), the problem
of computing the correct imagery for the viewer can be completely specified using two mappings.
The first mapping uses the geometric relationship among
the viewer, 3D model to be rendered, and the display surface
to determine the intensity values that must be assigned to
each point on the display surface. The second mapping uses
the exact relationship between each projector pixel and the
display surface point it illuminates in order to assign the
appropriate pixel intensities to each projectors’ frame buffer.
When the complete geometry is known exactly, the projections from all the frame buffers illuminate the scene so
that the viewer sees the correct imagery and the problem of
generating large-scale panoramic imagery is solvable. But
an understanding of the geometry is not enough to create a
practical system. Challenges arise when large-scale systems
are coupled with an inaccurate knowledge of geometry, incomplete or inaccurate camera/projector models, and when
performance becomes a concern.
Central to these geometric techniques is the use of cameras
to recover the 3D representation and calibrate the system.
The projectors and cameras operate together as a tightlycoupled system to recover critical geometric information of
the projectors and the entire display surface. The Office of
the Future system [7, 8] originally introduced the idea of using cameras to recover 3D displays surfaces and rendering
on them. We present new techniques that enable the system to maintain geometric registration and deliver seamless
panoramic imagery. Figure 1 shows the kinds of geometry
that our camera-based approach will allow: irregular display
surfaces, arbitrarily-positioned (and overlapping) light projectors, and a moving viewer. Our techniques do not require
idealized display surface geometry and viewer/projector configurations, although we encompass these special situations,
in some cases with improved efficiency.
The technical presentation begins with an application of
computer vision techniques to calibrate the cameras and recover display surface and projector parameters. We then
describe a two-pass rendering method for generating a perspectively correct image on an irregular surface. We then
detail issues with scaling the system: repeating these procedures for a large number of projectors. The individual steps
involved are direct, and yet unavoidable imperfections introduce errors in estimation of recovered parameters and lead
to visible artifacts in projector overlaps.
Our approach provides ways to compensate for the practical geometric errors that, left alone, lead to serious misregistration.
We introduce a surface mesh unification
method to ensure the geometric continuity of the 3D global
display surface model across projector boundaries. Finally,
to address the relatively small registration errors that inevitably remain after any parametric calibration, we present
a new 2D post-rendering warp step.
Geometric Relationships
The display surface and viewer physical environment defines
the WCS, and each projector and camera is modeled as a
perspective projection device in the WCS. The perspective
transformation relates the 2D pixel coordinates (image pixels
for the camera, frame buffer pixels for the projector) to the
3D WCS. Perspective transforms can be represented with a
pinhole camera model. The pinhole model includes intrinsic characteristics, such as the image center and the focal
length, as well as the extrinsic characteristics, which locate
the optical center and 3D orientation of the projective device.
Related Work
Although the idealized representation of the panoramic display problem can be stated quite directly, the implementation of a real system faces the practical issues of imprecise geometry, aliasing, blending and many sources of misregistration. There are a variety of ways to cope with these
issues, and many panoramic displays with and without user
head-tracking have been developed. The majority of the systems, such as those from Panoram Technologies [5] and Trimension Systems [4], create images for a single ideal viewer
location, or “sweet spot”. Specifically, Trimension [4] uses
three overlapping projectors to project images on a rigid
cylindrical screen. The light projectors are aligned symmetrically so that each overlap region is a well-defined rectangle
(Figure 2a). Flight simulators have been using a similar
technique for a long time. Omnimax [2] and ARC domes
[6] immerse the user in high resolution wide-field images
using a single projector and dome shaped surfaces. Using
rear-projection and head-tracking, the CAVE [1, 9] enables
interactive and rich panoramic visualizations. The setup is
a precise and well designed cube-like structure (Figure 2b).
The CAVE assumes that the display surface and projector
geometries are known and are fixed a priori in a specific
configuration. Geometric registration is obtained by carefully ensuring that the physical configuration matches the
The Office of the Future [7] suggests using arbitrary dayto-day surfaces for display purposes and rendering perspec-
tively correct images for a moving user on them. That system demonstrated ideas on a small-scale system, where the
idealized geometries and models are accurate enough to produce acceptable results. Our methods carry that small-scale
work forward, scaling to many cameras and projectors and
presenting new techniques that are necessary to obtain the
performance and the geometric registration required to deliver seamless panoramic display.
The issues of maintaining geometric registration between
projected images for a moving user have not been fully explored by current projector-based systems. However, many
authors have presented techniques to create panoramic mosaics of images taken with a camera. Typically, the images
are taken with a camera mounted on a rotating tripod. If
there is no strong motion parallax, the images are “stitched”
and smoothly blended to create a single panoramic image.
Earlier stitching methods required pure (horizontal) panning
motion of the camera [10, 11]. This is analogous to current
multi-projector systems that allow only side-by-side overlaps
and align two projectors at a time.
Newer panoramic image mosaicing techniques allow uncontrolled 3D camera rotations [12, 13] by representing each
image with a 3-parameter rotational model or sometimes
with more parameters. This allows mosaicing of images
taken with even a hand-held camera. We extend this concept and represent the image displayed by each light projector by a sequence of two perspective projection transformations. The panoramic imagery is created using arbitrary
projection overlaps. Most of the camera image mosaicing
techniques deal with the difficult problem of computing image feature correspondences. We reduce this problem by
using active structured light.
With regard to intensity blending of overlapping projected
images, some popular techniques are described in [14, 12].
Section 4.3 details how we modify similar techniques for our
multi-projector display environment.
As background for the discussion of multi-projector system
calibration in Section 4, this section details the three fundamental calibration steps for a single projector display and
includes a brief on the generalized 2-pass rendering technique
used as a basis of our work.
Display Surface
extrinsic parameters. These methods are based on standard computer vision techniques and systematically build
on each other. The camera pair and calibration pattern are
not needed during rendering operation and can be retired
once the calibration procedures are complete.
Camera Calibration
To calibrate the camera pair, we position a 3D calibration
pattern with spatially-known feature points within the intersection of their view frusta. By extracting feature points
in the 2D camera images corresponding to known 3D points
on the calibration pattern, we can determine the 3 × 4 projection matrix P̃ for each camera based on the perspective
[ U V S]T = P̃ [ x y z 1]T
The perspective equation maps a 3D point (x, y, z) in object space to a 2D point (u, v) in camera image space, where
(u, v) = ( U
S , S ). The projection matrix P̃ , determined up
to a scale factor, represents a concatenated definition of the
camera’s intrinsic and extrinsic parameters assuming a pinhole optics model. With six or more correspondences between the calibration pattern and camera images, the 11
unknown parameters of P̃ can be solved using a least-squares
method [15]. Note, we do not explicitly solve for the intrinsic and extrinsic parameters of the camera, instead we solve
for P̃ directly; however, the intrinsic and extrinsic can be
obtained by a decomposition of P̃ .
Display Surface Estimation
After independent calibration of each camera, we can evaluate the geometry of the display surface using triangulation techniques based on correspondences extracted from the
stereo image pair. Correspondences in the images are easily
determined since we can use the projector to sequentially
illuminate point after point until we have built a 3D point
cloud representing the display surface. By binary-coding the
projector illuminated pixels, we can efficiently determine the
stereo-correspondences. The process of projecting patterns
so that they can be uniquely identified by a camera is also
known as an active structured light technique.
This 3D surface representation, which is in the coordinate
frame of the physical calibration pattern established in step
one, is then reduced into a mesh structure in projector image
space using 2D Deluanay triangulation techniques.
Projector Calibration
Right Camera
Left Camera
Figure 3: Configuration for single projector.
Our calibration procedures are based on using a stereo
camera pair that is positioned on a wide baseline with each
oriented to observe the entire projector illuminated surface.
Figure 3 illustrates this layout. Step one of the procedures
involves calibration of the camera pair using a physical calibration pattern. Details of the physical calibration pattern are discussed in Section 6 and the accompanying video.
Calibration step two involves estimation of display surface
geometry and step three evaluates projector intrinsic and
As a result of the surface extraction process, for each projector pixel (u, v), we now have a corresponding illuminated
3D surface point (x, y, z). Using these correspondences we
can solve for the projector’s projection matrix P̃ as we did
for the cameras.
A problem arises when the 3D points of the display surface are co-planar. In this case, the least-square method is
degenerate due to the depth-scale ambiguity of viewing planar points. This means there exists a family of solutions.
To develop a unique solution in this case, we add surfaces
into the scene and repeat the surface extraction procedures
of Section 3.2 solely for the purpose of eliminating this ambiguity. Once this solution is affected, the introduced surfaces
are removed from the display environment.
2-Pass Rendering Algorithm
To render perspectively correct imagery on irregular surfaces, we use a two-pass rendering method described in [7].
In the first pass, the desired image for the user is computed
and stored as a texture map. In the second pass, the texture is effectively projected from the user’s viewpoint onto
the polygonal model of the display surface. The display surface model, with the desired image texture mapped onto
it, is then rendered from the projector’s viewpoint. This
is achieved in real-time using projective textures [16]. The
rendering cost of this two-pass method is independent of
complexity of the virtual model.
From a practical implementation standpoint, aliasing artifacts of the projective texture step can be reduced by computing a view frustum with the image plane parallel to the
best-fit plane representation of the display surface.
observes a set of projector pixels from projector Pi , but can
also observe a subset of the pixels projected by adjacent
projector Pi+1 . Similarly, camera pair Ci+1 observes pixels
projected by Pi+1 and a subset of the pixels of projector
Pi . The common set of 3D points that one camera pair
can observe from both projectors is the correspondence set
necessary to solve (2) and thus register the display surface
data of one projector to another.
Projection parameters of the projectors Pi are based on their
display surface Di as described in Section 3.3. After the
display surface meshes have been registered by applying rigid
transforms, we recalculate the projector’s projection matrix.
The remainder of this paper will address the issues in scaling
the calibration and rendering techniques from a single to a
multi-projector system.
First, to calibrate multiple projectors we repeat the procedures discussed in Section 3, but then we must re-register the
display surface definitions and projector calibrations for the
entire system to a common world coordinate space (WCS).
These registration methods are described in Sections 4.1
and 4.2, while Section 4.4 discusses re-registering the viewer
tracker data.
Second, display surface regions where multiple projectors
overlap are noticeably brighter because of multiple illumination. We correct for this by attenuating projector pixel
intensities in the overlapped regions. Our current intensity
blending technique are explained in Section 4.3.
Projector Registration
Projector Overlap Intensity Blending
Regions of the display surface that are illuminated by multiple projectors appear brighter, making the overlap regions
very noticeable to the user. To make the overlap appear
seamless we use alpha blending techniques. We create an
alpha-mask for each projector, which assigns an intensity
weight [0.0 − 1.0] for every pixel in the projector. Weights
of all projected pixels illuminating the same display surface
point should add up to unity. The weight is additionally
modified through a gamma lookup table to correct for projector non-linearities.
To find the alpha-mask, we use a camera to view the overlapped region of several projectors. We form a convex hull
Hi in the camera’s image plane of observed projector Pi ’s
pixels. The alpha-weight Am (u, v) associated with projector
Pm ’s pixel (u, v) is evaluated as follows:
αm (m, u, v)
Am (u, v) = P
α (m, u, v)
i i
Surface Mesh Registration
When multiple projectors Pi and stereo camera pairs Ci are
used, it is generally necessary to move the physical calibration pattern so that it can be viewed by the different camera
pairs. As described in Section 3.2, parameters for projector Pi and the corresponding section of the display surface
mesh Di are defined in the coordinate system of the calibration pattern used in the camera pair Ci calibration step. To
render seamless images, we first register all sections of the
display surface mesh into a common WCS.
Registering data represented in multiple coordinate
frames into a common frame is a classic computer vision and
graphics problem that involves solving for the rigid transformation given by:
Di (k) = Ri ∗ Di+1 (k) + t~i
where Ri is a 3 × 3 rotation matrix, t~i is a 3 × 1 translation
vector and Di (k) and Di+1 (k) are corresponding 3D points
in the two frames of reference. To compute R and ~t, we use
the Lagrangian multipliers method which solves the leastsquare minimization problem, ||Di (k) − (Ri Di+1 (k) + t~i )||2
subject to the constraint that Ri is a unitary matrix, i.e.
Ri RiT = I. This method is outlined nicely in [17].
The challenge in solving for R and ~t in most applications
is finding the correspondence between 3D points. We easily
find these corresponding points in Di+1 with Di using the
same binary-coded structured light methods used for surface
extraction and projector calibration. The camera pair Ci
where αi (m, u, v) = wi (m, u, v)∗di (m, u, v) and i is the index
of the projectors observed by the camera (including projector m).
In the above equation, wi (m, u, v) = 1 if the camera’s observed pixel of projector Pm ’s pixel (u, v) is inside the convex
hull Hi ; otherwise wi (m, u, v) = 0. The term di (m, u, v) is
the distance of the camera’s observed pixel of projector Pm ’s
pixel (u, v) to the nearest edge of Hi . Figure 4 shows the
alpha masks created for three overlapping projectors.
Projector 1
Projector 2
Projector 3
Figure 4: The top image shows the overlap position of three
projectors. The bottom images show the alpha masks created for projectors 1, 2, and 3 using our algorithm.
Similar feathering techniques are used to mosaic multiple
images taken by a camera into a single panoramic image
[18]. However, such a target 2D panoramic image does not
exist when projector images are blended on (possible nonplanar) display surfaces. Hence we use the 2D image of the
display surface taken from a calibration camera and compute
the intensity weights in this 2D camera image space. We
transformed this alpha-mask in camera space into projector
image space by using our 2-pass rendering technique. With
a fixed alpha-mask for each projector, we simply render a
textured rectangle with appropriate transparency as the last
stage of the real-time rendering process.
Tracked Viewer Registration
The user location is measured by an infrared tracking device. To use the tracker readings, we need to compute the
transformation between the coordinate system of the tracking system and the WCS defined by the physical calibration pattern. Therefore, before moving the physical calibration pattern from our reference projector, we also find the
correspondences between four 3D points on the calibration
pattern in calibration pattern coordinates and tracker system coordinates. We again use the Lagrangian multipliers
method to find the rigid transformation between tracker coordinate space and display WCS.
Errors in tracked user location cause no geometric registration problems between projectors as long as all rendering
processes are fully synchronized. However, errors in the estimation of the display surface and the camera/projector
parameters are critical. When these errors are small, the
display is seamless. When the errors grow, the display becomes discontinuous where projectors overlap. Methods for
addressing second-order errors such as for radial lens distortion and non-linear bundle adjustment to improve the
global estimate of the camera/projector calibration can reduce the average mis-registration error, but our experiments
have shown that radial distortion and bundle adjustment
improve estimates only marginally, and do not prevent the
viewer from perceiving mis-registrations in large-scale environments.
Therefore, to yield acceptable visual results we need to
address the two primary sources of error - display surface
estimation and projector calibration. We next review these
error terms and then present new calibration and rendering
techniques for preserving a seamless display.
Display Surface Errors
Display surface estimation is dependent on calibrated cameras and piecewise-registered 3D points that are connected
into meshes. Small errors that are present in camera calibration are magnified as errors in the final computed display
surface, resulting in an erroneous display surface. These
errors are attributed to the following operations: display
surface estimation include initial camera calibration from
the calibration pattern, which requires feature detection and
camera-parameter estimation; feature detection and stereo
correspondence from structured light techniques; recovery of
3D coordinates using triangulation; coordinate frame registration of meshes by estimating a rigid transformation to
bring them into alignment. Because the final display surface
will never be exact, we have developed techniques to lessen
the effect of display surface errors on the final rendered imagery.
Projector Calibration Error
Projector calibration is dependent on the 3D display surface
points that are reconstructed by the camera system. The relationship between the 3D points and the 2D pixel locations
that illuminate those points is the basis for calibration. Because of errors in the computed location of the 3D points, the
projection matrix for light projectors does not map the 3D
display surface points exactly onto their 2D pixel locations.
The re-projection error is directly dependent on errors in
the 3D display surface reconstruction; errors are small when
only one or two projectors and cameras are used, but grow
as the system scales. Due to these errors in the projection
matrices of light projectors, we have developed an imagebased technique to lessen the effect of re-projection errors
on the final rendered imagery.
Geometric Error Compensation
In order to create geometrically seamless imagery for the
viewer, we compensate for errors in 3D (display surface)
and 2D (calibration of projector in the form of a projection matrix from 3D display-space to the 2D frame buffer).
Specifically, we base our methods on two objectives:
• neighboring projectors should use exactly the same representation of the display surface geometry
• the projection matrix for a light projector should map
3D screen points onto the exact 2D pixels that illuminated them during the structured light process
If the overall environment reconstruction process was accurate, both objectives would automatically be satisfied. However, because it is inevitable that inaccuracies exist, our approach is to enforce geometric continuity in the registered
display surface in the projector overlap regions, and to guarantee geometric fidelity of the final imagery illuminated by
the projector. In the next two section we present two techniques for accomplish these goals.
Surface Mesh Unification
The objective is to create a single representation of the display surface from the multiple meshes recovered by different
stereo camera pairs. A single unified display surface will
not have discontinuities in regions where projectors overlap,
reducing geometric mis-registrations. The rigid transformation applied to each of the meshes brings them into near
alignment, but discontinuities still exist due to errors in the
3D recovery.
Specifically, two distinct but overlapping meshes are
brought into approximate alignment in a common coordinate system using the set of corresponding points that overlap between the two and are seen by the stereo camera pair
(described in Section 4.1). Stereo pairs Ci and Ci+1 may
both see illuminated pixels from projectors Pi and Pi+1 , and
such corresponding points are used for the alignment. After
the rigid transformation to align the two meshes, however,
3D values assigned to the illuminated surface points by Ci
and Ci+1 do not agree. Agreement is necessary, and we enforce it through a smooth 3D transition algorithm to obtain
display surface continuity.
Our present technique is similar to methods used to reduce intensity discontinuities in composited images [14][18].
However, instead of weighting pixel intensities, we weight
their associated 3D location. As with our intensity blending
algorithm, we use a single camera from a camera pair to aid
0 00
00 00
0 00
0 1111
0 00
11 11
00 11
11 11
11 11
1 1
0 00
11 00
11 1
D i+1
Figure 5: 3D points of the two display surfaces do not
agree after registration with a rigid transformation so we use
weighted averaging to obtain geometric continuity across the
surfaces. (see color plate).
with the averaging. The algorithm, which we term surface
mesh unification, works as follows:
Let Mi (u, v) be the 3D point associated with projector
P ’s pixel (u, v) seen by camera pair Ci . The new “weighted”
assignment of projector P ’s 3D point M (u, v) is evaluated
as follows:
M (u, v) =
Mj (u, v)dj (u, v)
dk (u, v)
where j and k are the index of cameras pairs that have
viewed this projector pixel. The term di (u, v) is the distance
of the observed projected pixel (u, v) of P to the nearest invisible (black) pixel defined in the in camera space Ci .
Using this weighted averaging technique, we obtain new
display surfaces Di0 that have geometric continuity. Note
that while the surface is continuous, it no longer represents
the true surface. We denote the modified surface points by
M 0 to distinguish them from the true surface points, M.
Figure 5 shows an overview of the algorithm for a simple
case of unifying two meshes. The same technique is used
when more than two meshes partially overlap.
Projector Coordinate
Figure 6: (a) Error in display surface estimation and projection matrix P1 creates mis-registration. (b) Partial correction using a post-rendering warp.
Using Di0 we now recompute the projection matrix for
the corresponding projector, as described in sections 3.3 and
4.2. The result is a new projection matrix correspondingly
denoted Pi0 . As shown in Figure 6 this new projection matrix maps the modified 3D surface points M 0 on Di0 to the
projector pixels m that illuminated them.
Post-rendering Warp
It is important to realize that because the transformation
from M to M 0 is non-rigid, the projection matrix Pi0 for each
projector cannot exactly map the points M 0 to m. Instead,
the projection matrix Pi0 maps the point M 0 to the distorted
location m00 . In this case for projectors 1 and 2,
m1 = P10 M 0
m2 = P20 M 0 .
What one would really like is the non-linear projective
function that directly maps M 0 to m. This function could
be determined by some other means, but the result could
not be implemented using the single linear projection matrix common in conventional graphics hardware. We achieve
this projective function in real time by first using P 0 as the
traditional linear projection matrix, and then following this
with a 2D post-rendering warp that maps m00 to m. The 2D
warp is based on a dense grid of sampled points from the
structured light process.
The texture-map implementation of this warp loads the
image generated with projection matrix Pi0 into texture
memory. The post-rendering warping is achieved using multiple textured triangles. This resultant image is projected
by the projector. All the warping operations, including 2pass projective texture rendering to create an image and
the 2D post-rendering warp, are fixed for given display surfaces, projection matrices and re-projection errors. Hence
they are established during pre-processing and loaded in a
display list. The cost of this post-rendering warp remains
fixed for given display surface and re-projection errors. It is
independent of the graphics virtual model being rendered.
There are two important notes to make. First, it is difficult to compute explicit projector pixel correspondences,
such as m1 and m2 . The correspondence is implicitly calculated by observing a dense grid of projected pixels. The
tessellated display surface geometry is simplified to improve
rendering speed during second pass of the two-pass rendering
method. We are investigating methods to simplify the number of triangles used during the post-rendering warp. Second, the projection matrices Pi0 that are actually computed
for each projector use the 3D surface points from the unified surface mesh Di0 , which have been unified as described
above. The computation of the unified surface mesh and
the post-render warp are done only once and can therefore
be implemented in real-time; the two techniques are closely
linked to one another.
Mesh unification and the 2D post-rendering warp meet
the two desired objectives: neighboring projectors should
use exactly the same representation of the display surface geometry, and that the projection matrix for a light projector
should map 3D screen points onto the exact 2D pixels that
illuminated them during the structured light process. By
applying these novel techniques we can guarantee seamless
geometric registration between overlapping projected images
even when the estimated display surface and projector parameters have large errors.
The system setup includes five 1024×768 resolution SHARP
LCD projectors and multiple JVC and Pulnix 640 × 480 resolution camera. The projectors are ceiling mounted approximately three to four meters from the display surfaces. These
projectors are casually positioned with multiple overlapping
regions to produce a 180 degree field of view when the user
is in the display center.
The calibration of the system (i.e., evaluation of camera
and projector parameters and display surface estimation) is
done once as a pre-rendering step. This is accomplished using a 0.6 meter cube that we constructed as our physical
target pattern and a Dell NT workstation equipped with
OpenGL graphics, Matrox Meteor II frame grabbers and
Matlab software. The equipment is first used to capture
the images of the physical target pattern and calibrate the
cameras. Next, the workstation performs the structured-
light projection and analysis, controlling one projector and
a stereo camera pair at a time. The stereo correspondences
acquired by projecting structured light form the dataset
needed for projector calibration, display surface reconstruction and unification, post-warp mesh generation, and alphamask generation. The actual processing for these steps is
done off-line using Matlab.
The required sampling density of the structure-light patterns depends on the complexity of the display surfaces and
the need to accurately locate the edges of overlapping projectors for alpha-mask generation. For our purposes, we used
sampling density of every 8th and every 32nd display pixel.
By binary encoding the structure-light, this process can be
parallelized and we are able to project and recover 16 × 12
correspondence points simultaneously. The complete operation for display surface recovery and light projector parameter estimation takes approximately 15 minutes per projector
at the highest sampling density and less than one minute for
the lower sampling density.
A moving user is tracked using an Origin Instruments’
DynaSight(TM) infrared tracker [19]. The user wears a set
of infrared LED beacons provided with the tracker. Tracker
readings are acquired, processed (low-pass filtered and transformed into the WCS) by a Dell NT workstation before being
dispatched in a network packet to the SGI image generation
The graphics rendering is done on an SGI InfiniteReality2
for each projector using the OpenGL API. While our rendering pipeline has additional computational cost due to the
image warping steps, this cost is fixed and is independent of
the rendered scene complexity.
Figure 7 shows a portion of our setup with three projectors forming a seamless panoramic image. The accompanying video shows the projection environment and real-time
operation on irregular and planar surfaces with a tracked
user. In addition, the video demonstrates how a non-expert
user can easily and quickly setup and use a two projector, head-tracked display. More images are available at
Figure 7: The top set of images show the individual contribution of three projectors. The bottom two images show
the projectors without and with alpha-blending. (see color
We have presented a general solution for creating a large
area display from many projectors where the display surface
can be irregular. The methods incorporate the case where a
head-tracked user views the display. The techniques for this
most general case can be simplified under certain conditions
such as when the viewer is static rather than moving, or
when the display surface is known to be planar.
Static User
Display systems with only a single “sweet spot” are commonly used because either the application guarantees that
the user will always stay in a single location (i.e., flight simulator) or that many people will view the images simultaneously from or near the correct position, as in domed displays such as the Omnimax [2]. The relationship between desired image and projected image for each projector, i.e., the
viewer-to-display mapping function, needs to be computed
only once and subsequently remains fixed for that location.
This mapping function can be obtained directly by using a camera to represent the viewer at a given location.
The camera observes points illuminated by projectors in
the display environment, establishing viewer-to-display correspondences. A detailed implementation of this method is
described in [20]. Using this technique, the rendering process has two stages; (1) compute the desired image and load
it into texture memory, and (2) warp the texture via the
viewer-to-display mapping to produce the correct imagery.
Intensity blending of overlapping projectors is handled as in
Section 4.3. This special case avoids explicitly solving for
3D parameters or the additional cost of the third-pass postrendering warp, but limits the user to one position in the
display environment.
Planar Surfaces
Surround screen displays can be easily created with multiple planar walls with a single projector illuminating a single
planar surface. Examples of such systems include the CAVE
[1, 9] and Trimension’s Reality Room[4]. The latter uses
cylindrical screens, but the section of the screen illuminated
by each projector is approximated by a plane.
In case of irregular surfaces, the warping function needed
in order to produce correct imagery must be expressed using a per-pixel mapping. In practice, this is accomplished
with the two-pass algorithm outlined in this paper. When
every projector illuminates a planar surface, a single-pass
rendering algorithm can achieve the same geometrically registered result at less computational expense. Specifically, the
planar constraint allows a single projector’s imagery to be
expressed as a collineation with the planar surface, which is
a 3 × 3 matrix that is updated as the user moves. Neighboring projectors are also related by a 3 × 3 collineation. The
special planar geometry allows for a rendering method that
does not incur additional computational costs while staying
geometrically registered [21]. The video shows results for
five projectors rendering onto two planar walls.
Although we have addressed how to remove many of the
geometric constraints involved in panoramic display systems,
this is only a basic framework. There are several important
issues that we have not fully addressed in this work that
require further research:
• arbitrarily large numbers of projectors
• automatic system calibration without the need for
a physical calibration pattern or human intervention
• error sensitivity analysis and the quantifiable impact on rendering accuracy
• detailed projector colormetrics and methods for
improving photometric seamlessness
• viewer-dependent photometrics including the issues of user location, surface orientation and intensity
distribution, and surface inter-reflections
• synchronization of rendering pipelines.
One can also extend our methods to display imagery for
multiple tracked users or to display stereo imagery by rendering for both eyes. In stereo displays, traditionally, left
and right eye images are projected by the same projector.
Using the geometric registration techniques, we can even use
two different sets of projectors (with different polarization,
for example) to render for left and right eye.
of our test environment, and Todd Gaul for video editing
support. A special thanks is also due Su Wen and 10-year
old Sam Fuchs for their participation in the video.
[1] C. Cruz-Neira, D. Sandin, and T. DeFanti, “Surround-screen
Projection-based Virtual Reality: The Design and Implementation of the CAVE,” Aug. 1993.
[2] N. Max, “SIGGRAPH 84 call for Omnimax films,” Computer
Graphics, vol. 16, pp. 208–214, Dec. 1982.
[3] K. Jarvis, “Real Time 60Hz Distortion Correction on a Silicon
Graphics IG,” Real Time Graphics, vol. 5, pp. 6–7, Feb. 1997.
[4] Trimension Systems Ltd.
[5] Panoram Technologies, Inc.
In this paper we have presented techniques for building a
scalable panoramic display device from multiple, casuallypositioned light projectors. We maintain geometric registration of the overall panoramic display; employing cameras
for recovering display surface and light projector geometries
rather than imposing geometric constraints into the overall
display setting. This integration of cameras into the initialization phase of the system and the techniques for maintaining registration in the face of errors leads to our primary
• Geometric registration and seamless imagery is produced over a wide range of geometric configurations
• Generalized display configurations including support for irregular display surfaces and a head-tracked,
moving viewer
• Rendering efficiencies in special cases such as for
a static user or planar display surfaces
• Post-rendering warp and 3D surface unification
together compensate for inaccuracies in the extracted
geometric representation
• Self-configurable display and projector geometries
are enabled so that future systems can automate calibration and registration.
Based on these new techniques, we believe we have built
and and demonstrated the very first panoramic display system using irregular surfaces for a moving user. Even though
there is much work yet to be done, this “proof-of-concept”
system previews future possibilities for high-resolution, wide
field-of-view displays that are easy to set-up, use and maintain.
[6] D. Bennett. Alternate Realities Corporation, Durham, NC
27703. Cited July 1999.
[7] R. Raskar, G. Welch, M. Cutts, A. Lake, L. Stesin, and H. Fuchs,
“The Office of the Future: A Unified Approach to Image-Based
Modeling and Spatially Immersive Displays,” in SIGGRAPH 98
Conference Proceedings, July 1998.
[8] R. Raskar, M. Cutts, G. Welch, and W. Stürzlinger, “Efficient
Image Generation for Multiprojector and Multisurface Displays,”
in Proceedings of the Ninth Eurographics Workshop on Rendering, (Vienna, Austria), June 1998.
[9] Pyramid Systems.
[10] E. Chen, “Quicktime VR - An Image-Based Approach to Virtual Environment Navigation,” in SIGGRAPH 95 Conference
Proceedings, pp. 29–38, Aug. 1995.
[11] L. McMillan and G. Bishop, “Plenoptic Modeling: An ImageBased Rendering System,” in SIGGRAPH 95 Conference Proceedings, pp. 39–46, Aug. 1995.
[12] R. Szeliski, “Video Mosaics for Virtual Environments,” IEEE
Computer Graphics and Applications, vol. 16, pp. 22–30, Mar.
[13] H. Sawhney and R. Kumar, “True Multi-Image Alignment and
its Applications to Mosaicing and Lens Distortion Correction,”
in IEEE Comp. Soc. Conference on Computer Vision and Pattern Recognition (CVPR’97), 1997.
[14] P. J. Burt and E. H. Adelson, “A Multiresolution Spline with
Applications to Image Mosaic,” ACM Trans. on Graphics, no. 2,
pp. 217–236, 1983.
[15] O. Faugeras, Three-Dimensional Computer Vision: A Geometric Viewpoint. Cambridge, Massachusetts: MIT Press, 1993.
[16] M. Segal, C. Korobkin, R. Widenfelt, J. Foran, and P. Haeberli,
“Fast Shadows and Lighting Effects using Texture Mapping,” in
SIGGRAPH 92 Conference Proceedings, July 1992.
[17] R. Haralick and L. Shapiro, Computer and Robot Vision, vol. 2,
ch. 14. Addison-Wesley, 1993.
[18] H. Shum and R. Szeliski, “Panoramic Image Mosaics,” Tech.
Rep. MSR-TR-97-23, Microsoft Research, 1997.
[19] Origin Instruments Corporation.
This research is supported by the National Science Foundation agreement ASC-8920219: “Science and Technology
Center for Computer Graphics and Scientific Visualization”,
Link Foundation, Intel Corporation, and the “National TeleImmersion Initiative” sponsored by Advanced Networks &
Services, Inc.
We would like to thank Gopi Meenakshisundaram, Aditi
Majumder and David Marshburn for useful discussions and
support. We also gratefully acknowledge John Thomas, Jim
Mahaney and David Harrison in the design and assembly
[20] R. Raskar, G. Welch, and H. Fuchs, “Seamless Projection Overlaps Using Image Warping and Intensity Blending,” in Fourth
International Conference on Virtual Systems and Multimedia,
(Gifu, Japan), Nov. 1998.
[21] R. Raskar. Olique Projector Rendering on Planar Surfaces for a
Tracked User.
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