A PARALLEL MEDIATED REALITY PLATFORM
Rosco Hill, James Fung and Steve Mann
University of Toronto
10 King's College Road
Realtime image processing provides a general framework for ro-
bust mediated reality problems. This paper presents a realtime
mediated reality system that is built upon realtime image process-
ing algorithms. It has been shown that the graphics processing
unit (GPU) is capable of efficiently performing image processing
tasks. The system presented uses a parallel GPU architecture for
image processing that enables realtime mediated reality. Our im-
plementation has many benefits; the graphics hardware has high
throughput and low latency; the GPU's are not prone to jitter. Ad-
ditionally, the CPU is kept available for user applications. The
system is easily constructed, consisting of readily available com-
A "mediated reality system" (MRS) is one that can augment, di-
minish, or otherwise alter the visual perception of reality. By way
of explanation, "virtual reality" creates a completely computer-
generated environment, "augmented reality" uses the existing real-
life environment and adds computer-generated information thereto,
"mediated reality" filters and adds to reality to enhance the visual
information presented to the user.
A tetherless mediated reality system typically runs on a lower
power battery-powered wearable [1, 2] computer system with minia-
ture eyeglass mounted screen and appropriate optics to form the
virtual image equivalent to an ordinary computer desktop. How-
ever, because the apparatus is tetherless, it travels with the user.
Users wearing a MRS could use it for information selection and
retrieval, personal safety, enhanced situational awareness and nav-
This paper presents a parallel GPU architecture that forms the
basis of a realtime MRS. There are three problems in mediated
reality that are addressed in this paper: world registration, ob-
ject recognition and scene generation. All three problems are ad-
dressed within the context of the parallel graphics architecture and
the solutions proposed run on the GPU's. Two proof-of-concept
applications are presented here to demonstrate the effectiveness
of parallel graphics architecture for realtimemediated reality. In
keeping with the parallel spirit of the MRS, the applications lever-
age the parallel GPU architecture.
The reality window manager (RWM) application is a user in-
terface for MRS. Traditional models for displaying information on
Thanks to NSERC and Ontario Graduate Scholarships for support.
Thanks to nVIDIA for equipment donations.
a computer display are constrained to a 2D coordinate systems.
For example the X11 window system. In some augmented reality
systems, windows simply float in front of the user. We present a
natural mapping that displays 2D information in a 3D world, sim-
ilar to the manner in which regular non-electronic information is
presented (plastered on signs and billboards). The RWM mediates
the contents of planar patches in the real world with information or
application windows such as terminals, web browsers, and email.
The RWM interface projects information more naturally than the
current WIMP standard. Figure 1 shows the author's desktop run-
ning the RWM application.
. The GL terminal application and the glynx web browser
registered on quadrilaterals in the author's environment.
The virtual architect application superimposes the model of a
building on top of the real building, which matches the perspective
of the viewer. This enables an architect to envision and commu-
nicate the look and feel of the final building. For example the Art
Gallery of Ontario recently unveiled the $180-million transforma-
tion designed by architect Frank Gehry. A video capture of the
model on display is used to overlay the current building so that the
viewer can see what the building will look like after construction
is finished in 2007.
Previous work on markerless tracking and registration typically
does not run in real time [3, 4]. In this paper we leverage the fea-
tureless realtime projective image registration system using paral-
lel graphics processing units that is described in .
Other researchers have presented realtime systems that use
specialized hardware or markers for regristration. Feiner et al.
 presented a system with realtime world registration and scene
synthesis using 3D trackers so that the real and virtual worlds
could be registered. This research explored three ways to present
augmented large windows that the user could explore by looking
around: floating windows, windows that remained fixed to a loca-
tion in the user's display, and world fixed windows. The system
presented by the author is much more flexible, it requires no spe-
cialized hardware and allows information to be blended onto any
planar surface in the environment. Azuma et al.  explored a
hybrid approach using inertial, optical and compas inputs to track
registration. In constrast, our system focuses on running on hard-
ware available at the local computer store.
It has been shown that graphics processing units are capable of
efficiently performing computer vision tasks [8, 9, 5]. The GPU's
found in most personal computers and laptops typically exceed, in
number of transistors as well as in compute power, the capabilities
of the CPU. Fung et al.  demonstrated a system that achived re-
altime video framerate projective image registration using a single
GPU. Previous work has discussed implementations which use a
single graphics card for realtime computer vision. Work has also
been done to apply GPUs to general purpose computing and other
specialized tasks beyond computer graphics alone
. Our focus is
to investigate how graphics cards can function in parallel as the
basis for a realtime mediated reality system.
The MRS has three main subsystems: world registration, object
recognition and scene generation. Each subsystem runs on the par-
allel graphics processing architecture.
The parallel graphics processing architecture consists of six
commodity graphics cards. Five cards are on the PCI bus and each
have a GeForce FX 5200 GPU. The sixth graphics card is on the
AGP bus and has a GeForce FX 5900 GPU. The subsystems dis-
tribute their workload amongst the graphics cards according to a
workload estimation heuristic. Figure 2 illustrates the architecture.
. Parallel GPU Architecture. 6 PCI graphics cards commu-
nicate with a faster AGP card via the PCI and AGP buses.
4. PARALLEL WORLD REGISTRATION
To effectively implement mediated reality in arbitrary environments,
the world registration must be accurate, fast and reliable. We
present afast and reliable image registration algorithm that forms
the basis for world registration in our realtime mediated reality
4.1. Image Registration
The system presented uses a parallelized version of the projective
image registration presented in  to achieve realtime robust reg-
istration. The mediated reality system incorporates a camera, with
a diverter material, such that the lens of the camera captures the
image that is going into the eye. The system tracks the relative and
absolute position and orientation of the camera in terms of the pro-
jective coordinate transformation (PCT) between the current frame
of video and the previous frame. This featureless image motion
estimation algorithm tracks the ego-motion of the camera so that
the virtual world remains aligned. The parallel algorithm solves
the Equation 1 to find the projection of an image to register itself
with another image.
The number of iterations required for the solution to converge
is related to the amount of change between two successive frames.
For example, two frames with little relative movement converge
in around 10 iterations, however, a sudden jerky movement of the
camera a laggy system cause the number of required iterations to
increase to as many as 50 (system delay and lag are discussed in
Section 8). The workload estimation heuristic distributes the work
amongst more cards according to the estimated number of itera-
tions required to converge.
4.2. Registered Mediated Reality
As the user scans his/her environment by looking around, the sys-
tem tracks the ego-motion of the head by solving successively for
the PCT. Images from the current camera orientations can be
synthesized by composing the relative PCT with the spatial loca-
tions of the virtual objects. In OpenGL this corresponds to apply-
ing the appropriate matrix multiplication to the model view matrix.
To illustrate this, consider Figure 3. We see a approximate bound-
ing planar patch that is being tracked by the system. As the user
looks to the left, the majority of the planar patch is outside of the
field of view of the user. The successive compositions of the PCT
obtained from the image registration algorithm keep the position
of the bounding planar patch registered with the real world.
5. OBJECT RECOGNITION
Our object recognition subsystem has two methods for detecting
objects. The first is quadrilateral contour detection for finding of
planar patches in the field of view. Information is typically pre-
sented on some 2D surface, and as many applications are con-
strained to a 2D coordinate system, we find planar patches to iso-
late the areas where information mediation will take place.
Figure 2: Outdoors: the environment map generated by wearer A, and the views of that map synthesized by wearer B.
Figure 3: Indoors: the environment map that wearer B generates while browsing wearer A's outdoor map.
use. Each image in the database includes a time stamp, a
3D rotation matrix describing the camera orientation and a
PCT describing its projective position with respect to the
base frame. Reference frames are selected for use with
VideoOrbits based on the current position estimate gener-
ated by the gyroscope.
The sphere of reference images is mapped to a two di-
mensional matrix, indexed by equal increments of azimuth
and elevation. In experimentation, four degree increments
3. Seeing Eye to Eye: a shared mediated reality
Sharing of the mediated reality is accomplished by trans-
ferring this reference frame database to wearer B (see Fig-
ure 1), who can use a second VOGHT and a browser pro-
gram to synthesize views of wearer A's environment map
from an arbitrary viewing direction. Synthesis of views is
performed by taking the current estimate of the rotational
position of wearer B's VOGHT and using its equivalent
PCT to reproject wearer A's reference images to wearer
The images that are projected are selected by the abso-
lute position of the VOGHT. This estimate is used to cal-
culate a region in which to apply the projective coordinate
transform. Then using the projected images an image com-
posite is formed. From this composite the correct view is
cropped (see Figure 2).
Thus the head tracking of wearer B is used to provide a
perspective with which to view the environment map gener-
ated by wearer A. As wearer B rotates his/her head, new
perspectives are supplied to the browser program, which
cause the projection and synthesis of new portions of the en-
vironment map. All these projections are relative to wearer
B's current head position, thus all navigation is performed
solely with head movement.
The size of the region to search is a tunable parameter,
a small parameter is useful on a resource constrained wear-
able computing system. If the search area is large then mul-
tiple images from the map can be projectively transformed
and registered (see figure 2). This registered image allows
for the creation of a complete image for the desired per-
spective even if there is no image in the map at that exact
perspective. However, if there are no images nearby then no
view can be generated.
The shared mediated reality system allowed users to ex-
change their current environments through the EyeTap re-
ality mediator. The views are generated and controlled
through the users' head motion using the VOGHT. This was
demonstrated through the use of the system between two
wearable computer users, one outdoors and one indoors.
The views generated had the correct perspective and al-
lowed each viewer to see the other's environment as if it
were their own. This resulted in the users seeing Eye to Eye
in a projectively stabilized shared mediated reality.
 C. Aimone, A. Marjan, and S. Mann. Eyetap video-based pro-
jective motion estimation assisted by gyroscopic tracking. To
appear in International Symposium on Wearable Computing,
 S. Mann. Intelligent Image Processing. John Wiley and Sons,
November 2 2001. ISBN: 0-471-40637-6.
Proceedings of the International Symposium on Mixed and Augmented Reality (ISMAR'02)
0-7695-1781-1/02 $17.00 © 2002
. Head Tracking: The virtual square remains aligned as the
user looks around.
Finding candidate quadrilaterals in a scene amounts to finding
either dark quadrilaterals on a lighter background or light quadri-
laterals on a darker background. The steps for the algorithm are
described below. The steps are repetitive in that they are repeated
for a number of scales in the image, and for a number of light-
The video image is pre-blurred with a gaussian filter to re-
move noise, then a multi-scale pyramid is formed. At each image
resolution we threshold the image and compute the approximate
quadrilateral bounding box for the detected contours. There is a
parallelism inherent in this computer vision algorithm, and each
video card runs the quadrilateral detector in parallel at different
thresholds. The result is communicated back to the MRS and the
best quadrilateral is selected for mediation.
The object recognition also has a template matching search
method based on Eigenspace techniques. In advance, many
images of an object are taken under various poses. The GPU
computes the basis images for the objects, and stores them in a
database. Once the database is populated, the template matcher
must search for the object under a variety of orientations and scales.
The computational requirements for traversing the search space for
the template matching algorithm is:
[(640 x 480) x 3 colors] x 12 orientations x 5 scales
The template matching algorithms typically search at 5 different
scales and each search is independent, hence the search can be
performed in parallel on different GPU's. Each GPU performs
the search at one scale, and communicates the result back to the
program. The program them decides if there is a match and the
reality mediator subsystem can diminish or augment information
on or around the object.
6. SCENE GENERATION
The MRS generates a scene that contains augmented or diminished
reality. The unmediated portions of the field of view are simply
re-rendered. The mediated portions are rendered as semi-opaque
displays on top of the original view. The RWM application ren-
ders a OpenGL based Unix
terminal window that matches the
position and orientation of the planar patch it is attached to. The
virtual architect application renders the model of a building on top
of the actual building, such that two registration points (i.e. the
front doors) remain aligned. It is important to note that the regis-
tration points do not need to be in every scene, only in the initial
scene. The user is free to look around the world and return his/her
gaze to the mediation zone and the mediated objects remain fixed
to their respective objects.
All scene generation is performed on the AGP graphics card.
The scene consists only of texture mapped quadrilateral, which
contain either the video signal or the information for augmenting
or diminishing reality. For the RWM application an OpenGL ter-
minal is rendered at a specified position, zoom and orientation and
for the virtual architect application a texture mapped quadrilateral
7. HUMAN-COMPUTER INTERACTION
Our implemenation allows users to specify a planar patch in their
field of view by either selecting 4 corners manually, or allowing
the object detection subsystem to automatically select the best pla-
nar patch from the field of view. By clicking the moues button,
the user attaches an application window to the object. The sys-
tem tracks the affine transformations from the ego-motion of the
camera and applies this transformation to all of the windows dis-
played on the desktop. The net effect is an application window
that remains registered on the planar patch in location, zoom, ori-
entation, keystoning and chirping, as shown in Figure 1. The user
is free to move and look around, and the application will remain
attached to that object when the user returns.
8. SYSTEM DELAY AND JITTER
Dynamic errors occur in tracking and registration because of sys-
tem delays. The total delay in an MRS system is the time from
moment the system captures an image until the time that the im-
age is re-synthesized to the user. MRS systems can include delays
from the tracking subsystem, communication delays, network de-
lays, scene recognition delays, scene generation delays and finally
scanout time from the frame buffer.
System delays cause error when motion occurs. Take the case
with a static scene with a virtual window attached to a billboard
shown in Figure 1. Suppose the user turns their head at time t
and the system delay is 100ms, then the user will see the virtual
window drawn at the same location in his/her field of view until it
is finally updated at time t + 100ms. These images are therefore
incorrect for the time when they are viewed and the two worlds are
The parallel MRS runs end-to-end on the graphics cards. The
frame-grabber data is transfered directly to the texture memory
on the GPU, and the GPU program runs as soon as the texture
memory is updated. Furthermore, the mixing of the physcial world
and the virtual world is performed on the video card. For a fixed
number of virtual windows, both aspects of the RWM execute in
a predictable time constant. This eliminates "jitter" and allows the
user to adapt to a constant system delay.
The system delay is extremely short on modern graphics cards.
Our prototype system has a NVIDIA NV35 chip, which is the GPU
found on the GeForce FX 5900 Ultra board. The NV35 GPU has
a transistor count of about 135 million. The system delay is 33 ms
for capture, 3ms per registration estimation, and 10ms rendering,
plus 16.67ms. Thus our system has an end-to-end lag of 46ms, or
roughly 1.5 frames on a 30fps signal. Our system lag is dominated
by the slow 30fps frame capture we use and a 60fps capture board
would cut this in half (though would likely increase the cost, size
and power consumption of the wearable).
It is believed that system delays are not likley to completely
disappear anytime soon. The most optimistic prediction made in
 is that HMD system with a 10ms lag might become possi-
ble in the next few years, but the drastic cut in throughput and the
expense required to construct the system would make alternate so-
lutions attractive. The paper also concludes that reducing end-to-
end delay to the point where it is no longer a source of registration
error is not practical. The system we present has no expensive cus-
tom parts. It runs in realtime at 30 frames per second. The signal
processing delay is 10ms, plus scan-in and scan-out time, while
costing only a few hundred dollars to build. Furthermore our sys-
tem is robust to registration errors.
This paper presents the first realtime markerless mediated reality
system using low-cost off the shelf computer graphics hardware.
The low cost and widespread availability of computer graphics
hardware will make hardware accelerated mediated reality pos-
sible. Additionally, the algorithms used in mediated reality are
inherently parallelizable. A parallel graphics architecture is pro-
posed and the delay and jitter of the system are analyzed. The
signal processing delay is under 45 ms and our system is robust to
registration errors. Two proof-of-concept mediated reality appli-
cation demonstrate the viability of the parallel graphics hardware
architecture. Both applications run within the parallel graphics ar-
chitecture and the CPU is kept available for other applications.
 S. Mann, "`mediated reality'," TR 260, M.I.T. M.L. vismod,
 Steve Mann, Intelligent Image Processing, John Wiley and
Sons, November 2 2001, ISBN: 0-471-40637-6.
 A.J. Azarbayejani, B. Horowitz, and A. Pentland, "Recursive
estimation of structure and motion using relaitve orientation
constraints," in Proc. CVRP, 1993, pp. 294299.
 K.N. Kutulakos and J.R. Vallino,
"Calibration free aug-
mented reality," in IEEE Trans. on Visualization and Com-
puter Graphics, 1998, pp. 120.
 James Fung and Steve Mann, "Computer vision signal pro-
cessing on graphics processing units," in To appear in the
Proceedings of the IEEE International Conference on Acous-
tics, Speech, and Signal Processing (ICASSP 2004), Mon-
treal, Quebec, Canada, May 1721 2004.
 Steven Feiner, Blair MacIntyre, Marcus Haupt, and Eliot
Solomon, "Windows on the world: 2d windows for 3d aug-
mented reality," in ACM Symposium on User Interface Soft-
ware and Technology, 1993, pp. 145155.
 Ronald Azuma et al., "Tracking in unprepared environments
for augmented reality systems," Computers and Graphics,
vol. 23, no. 6, pp. 787793, 1999.
 James Fung, Felix Tang, and Steve Mann, "Mediated real-
ity using computer graphics hardware for computer vision,"
in Proceedings of the International Symposium on Wearable
Computing 2002 (ISWC2002), Seattle, Washington, USA,
Oct. 7 10 2002, pp. 8389.
 Ruigang Yang and Pollefeys M., "Multi-resolution real-time
stereo on commodity graphics hardware," Computer Vision
and Pattern Recognition, 2003. Proceedings. 2003 IEEE,
vol. 1, pp. 211217, 2003.
 Felix Tang, Chris Aimone, James Fung, Andrej Marjan, and
Steve Mann, "Seeing eye to eye: a shared mediated real-
ity using eyetap devices and the videoorbits gyroscopic head
in Proceedings of the International Symposium
on Mixed and Augmented Reality (ISMAR2002), Darmstadt,
Germany, Sep 1 - Oct 1 2002, pp. 267268.
 M.J. Black and A.D. Jepson, "Eigentracking: Robust match-
ing and tracking of articulated objects using a view-based
representation," Proc. 4th European Conf. on Computer Vi-
sion, pp. 329342, April 1996.
 R. Azuma, "A survey of augmented reality," 1995.