books (IEBooks) for
surgical training will
let surgeons explore
procedures in 3D.
The authors describe
the techniques and
tools for creating a
embodying some of
the basic concepts.
geon-in-training’s ability to learn them. In
trauma-related procedures, for example, surgeons
must quickly stabilize the patient, and teaching
is secondary to patient care. Such procedures are
unplanned, and some critical traumas are rare.
This issue is not trivial. For most surgical spe-
cialties, physicians require at least five years of
postmedical school training before they are eli-
gible for board certification, and it is frequently
many years after board certification before most
surgeons are entirely comfortable with their craft.
In addition, as in many other technology-driven
fields, practicing surgeons must keep up-to-date
with a dizzying array of new devices and proce-
dures designed to improve care. Combined with
an overall increased focus on limiting medical
errors is an intense interest in improving and
maintaining surgical competency. Currently,
there is no educational environment that can
urgical training has traditionally fol-
lowed the adage, “See one, do one,
teach one,” but the complexity and rar-
ity of some procedures limits a sur-
replace the apprenticeship environment of the
operating room. Schools and hospitals can use
some teaching modules to teach and assess vari-
ous technical skills, but there is nothing similar
to the flight simulator for pilot training that can
replace being in the operating room (see the
“Visualizing Procedures in 3D” sidebar).
In 1994, Fuchs et al. outlined a vision and
some methods for using a “sea of cameras” to
achieve 3D telepresence for remote medical pro-
cedures.1We have since pursued related work for-
mally and informally with multiple collaborators.
In 2002, we presented an intermediate summary
of results from a three-year effort incorporating
teleimmersion technologies into surgical train-
ing.2Our goal is to create technologies that let
surgeons witness and explore a previous surgical
procedure in 3D as if they were present at the
place and time of the actual event, with the fol-
lowing added benefits: nonlinear control over
time; instruction from the original surgeon or
other instructor; and integrated 3D illustrations,
annotations, and relevant medical metadata. This
would let surgeons integrate the technical aspects
of the procedure with the anatomy, radiography,
and physiology of the patient’s condition, thus
best simulating and in many ways augmenting
the real-time educational experience of learning
how to perform an operative procedure. Figure 1
shows an early artist’s sketch of our immersive
electronic book (IEBook) and a screen shot from
our current prototype.
Here we present a current look at our efforts,
describing a set of tools that together comprise a
system for creating and viewing an IEBook and
displaying some results.
An understanding of the physical apparatus
involved should help readers better understand
how the system lets users create and view content.
The primary piece of acquisition equipment is
the camera cube shown in Figure 2. The unit is
constructed from rigid modular aluminum fram-
ing manufactured by 80/20 (http://www.8020.
net). Each side is approximately one meter.
On the unit are four fluorescent high-frequen-
cy linear lights from Edmund Industrial Optics,
and eight 640 × 480 Dragonfly FireWire (IEEE
1394) color cameras from Point Grey Research
(PGR). The lights reduce specular reflections and
provide even illumination. The cameras support
1070-986X/05/$20.00 © 2005 IEEE Published by the IEEE Computer Society
Greg Welch, Andrei State, Adrian Ilie, Kok-Lim Low,
Anselmo Lastra, Bruce Cairns, Herman Towles, and
University of North Carolina at Chapel Hill
University of Kentucky
Sascha Becker, Daniel Russo, Jesse Funaro, and
Andries van Dam
used with a PGR sync
unit, which ensures
that all eight cameras
eliminating the need
for temporal rectifica-
tion prior to recon-
struction of dynamic
3D models from the 2D
images, as we describe
in the next section.
the FireWire band-
width for 15-frame-
cameras into four cam-
era-pair groups. Each
The most obvious substitute for being in the operating
roomconventional 2D videohas long been available, but
surgeons universally consider these videos marginally effective
at best for several reasons:
Video relegates the viewer to the role of passive observer
who cannot interactively change the view.
❚ Subtle, complex motions that can be critical in an operation
are easy to miss.
❚ Video provides few depth cues and offers only linear control
over the playback’s timing.
Replaying video only provides the same experience a second
time, rather than permitting a new or changed perspective.
In short, watching a video is not much like experiencing the
procedure. Video is data from the procedure, but as Davis
points out, “experiences are not data.”1Davis describes expe-
rience as “an intangible process of interaction among humans
and the world that has its existence in human minds,” noting
that we can only capture and archive the materials (data) that
“occasion experiences in human minds.”1
Three-dimensional computer graphics systems let sighted
people see real or virtual models with the depth, occlusion rela-
tionships, and motion parallax they are used to in everyday life.
When coupled with interaction devices and techniques, such
systems can indeed occasion personal experiences in human
minds. Several disciplines have leveraged the power of this par-
adigm in training for many years. Arguably the most successful
example is flight simulation for pilot training.2Today flight sim-
ulators are considered so effective (and cost-effective) that pilots
sometimes use them to train on a new version of a plane, and
then fly the real plane for the first time on a regularly scheduled
flight with regular passengers.
Medical training and education programs have also used 3D
graphics techniques.3–5Previous training efforts have primarily
combined realistic models with interaction devices (sometimes
haptic) to simulate the experience of performing a particular
procedure. Rather than aiming to simulate the “do one” of the
adage, “see one, do one, teach one,” this work aims to improve
the opportunities to “see one,” and improve the associated
learning experience’s effectiveness.
1. M. Davis, “Theoretical Foundations for Experiential Systems
Design,” Proc. ACM SIG MultiMedia 2003 Workshop on Experiential
Telepresence (ETP 2003), ACM Press, 2003, pp. 45-52.
2. B.J. Schachter, “Computer Image Generation for Flight
Simulation,” IEEE Computer Graphics and Applications, vol. 1, no.
4, Oct. 1981, pp. 29-68.
3. J.F. Brinkley, “The Digital Anatomist Project,” Proc. Am. Assoc.
Advancement of Science Ann. Meeting, AAAS Press, 1994, p. 6.
4. R. Zajtchuk and R.M. Satava, “Medical Applications of Virtual
Reality,” Comm. ACM, vol. 40, no. 9, 1997, pp. 63-64.
5. A. Liu et al., “A Survey of Surgical Simulation: Applications,
Technology, and Education,” Presence: Teleoperators and Virtual
Environments, vol. 12, no. 6, 2003, pp. 599-614.
Visualizing Procedures in 3D
Figure 1. Immersive electronic book
(IEBook) in surgical training:
(a) early conceptual sketch of a blunt
liver trauma surgery, (b) the
acquired 3D scene of the surgery with
annotations and trainee’s hands
(sketches by Andrei State); and
(c) an IEBook in our present
immersive display environment. The
third of six annotated snapshots is
selected and playing. The red box
magnifies the snapshot for the fourth
group includes two PGR cameras, a FireWire hub,
a PGR sync unit, and a Dell PowerEdge rack-
mounted server running Microsoft Windows
2000 Professional. Figure 3 depicts the FireWire
connections within a group and between groups.
For reconstructing 3D models from the 2D
images, we use a standard personal computer
running Microsoft Windows XP Professional. We
currently use a Dell Precision 340 with a 2.2-GHz
Pentium 4 and 2 Gbytes of RAM.
Both content authoring and viewing take
place in a Barco TAN VR-Cube located at Brown
University. The VR-Cube, which is approximate-
ly three meters on each side, has four projection
surfaces (three walls and a floor) and four corre-
sponding Marquee 9500LC projectors configured
for a 1,024 × 768 screen with 60-Hz field-sequen-
tial stereo. The system uses five commodity
Linux workstations: one to gather device data
and four to render. The rendering workstations
are dual 2-GHz Intel Xeons with genlocked
nVidia Quadro FX 3000G graphics cards.
Myricom Myrinet connects the workstations for
high-speed communication. We use CrystalEyes
active liquid crystal shutter glasses, and
Polhemus FasTrak and
Intersense IS900 track-
The primary auth-
oring input device is a
handheld Tablet PC.
We currently use a
Toshiba Portege M200
to navigate the dynam-
ic 3D points/meshes in
time, initiate snap-
shots, highlight and
annotate, and arrange snapshots hierarchically.
Our primary system for fully immersive inter-
action with a complete IEBook is a combination
of the Barco TAN VR-Cube projection system and
the Acer Tablet PC. This equipment lets users
enjoy both full immersion and the complete set
of IEBook features available. This is currently the
only viewing option providing a complete expe-
We also explored a hybrid display system
combining head-mounted and projector-based
displays to simultaneously provide a high-quali-
ty stereoscopic view of the procedure and a
lower-quality monoscopic view for peripheral
awareness (see the “Visualizing Procedures Up
Close and All Around” sidebar).
We also support a relatively limited but more
accessible Web-based paradigm. The paradigm is
limited in that it provides reduced or no immer-
sion, and we currently only show dynamic 3D
points/meshes. We use VRML 97 to support
Web-based viewing of the dynamic 3D
points/meshes with viewpoint control. The
tions to implement the user interface actions.
The functions are compatible with VRML plug-
ins such as the Cortona plug-in (available at
Figure 2. Content
acquisition: (a) camera
cube used for
acquisition with a
medical training model
in place, and (b) a
performing a mock
procedure to manage
severe blunt liver
Figure 3. A single camera-pair group,
and the FireWire intra- and
intergroup connections. The four
groups are connected from previous
We generate anaglyphic (red-blue) stereo
images and movies from the dynamic 3D
points/meshes so users can view the data in stereo,
using an inexpensive pair of red-blue 3D glasses.
We select representative moments within each
sequence and render stereoscopic still images of
the points/meshes at those moments. We then
make these available with the Java Stereoscope
applet (see http://www.stereofoto.de). We also cre-
ate anaglyphic stereo movies in QuickTime and
Figure 4 gives an overview of the two major
phases involved in IEBook production: content
creation and content viewing. During content cre-
ation, the system acquires 2D video of an actual
Clearly, a student trying to understand a new surgical pro-
cedure would appreciate a high-quality, up-close view of the
procedure. But what is not so obvious is that learning how to
interact and deal with events and confusion around the patient
and in the room is also critical. A display with high-fidelity stereo
imagery at arm’s length and a wide (surrounding) view field can
engender this complete experience. Head-mounted displays
(HMDs) can provide the former, and projector-based displays
the latter, but unfortunately we are not aware of any display
system that simultaneously meets both needs.
We merged the two paradigms and developed a hybrid dis-
play system, shown in Figure A, that combines head-mounted
and projector-based displays to simultaneously provide a high-
quality stereoscopic view of the procedure and a lower-quality
monoscopic view for peripheral awareness.1We use a Kaiser
ProView30 stereo HMD, which does not have baffling material
around the displays that would block the wearer’s peripheral
view of the real world, and four projectors to render view-
dependent monoscopic imagery on Gatorboardand Styrofoam
arranged to match the operating table and nearby walls.
We evaluated the display’s efficacy with a formal human
subject study. We asked the subjects to solve a series of word
puzzles directly in front of them while simultaneously detect-
ing and identifying objects entering and exiting the periphery.
Our study indicated that peripheral distraction and associated
head turn rates were significantly less when solving both sets of
tasks using the hybrid display than an HMD alone.1
1. A. Ilie et al., “Combining Head-Mounted and Projector-Based
Displays for Surgical Training,” Presence: Teleoperators and
Virtual Environments, vol. 13, no. 2, Apr. 2004, pp. 128-145.
Figure A. Our hybrid display system combining head-mounted
and projector-based displays.
Visualizing Procedures Up Close and All Around
Content creationContent viewing
Figure 4. Overview of the phases involved in creating and viewing an immersive electronic book (IEBook). (RGB=Red,
Green, Blue; VRML=Virtual Reality Modeling Language; and WWW=World Wide Web.)
surgical event from multiple cameras, recon-
structs 3D models, and authors an IEBook. Figure
5 illustrates these processes.
The first step in creating an IEBook is captur-
ing (acquiring) the event of interest using the
camera cube. The basic process, depicted in
Figure 5a, involves calibrating the cameras, cap-
turing synchronized video, and converting the
raw images into the RGB color space for recon-
Camera calibration. To perform the vision-
based 3D reconstruction (see Figure 5b), we need
estimates of analytical models for the cameras’
geometric and photometric properties. We per-
form respective calibration procedures to prepare
for event capture.
For geometric calibration, we use Bouguet’s
Camera Calibration Toolbox for Matlab.3We take
a sequence of frame sets (sets of eight synchro-
nized images—one from each camera on the
camera cube) of a moving black-and-white
checked calibration pattern (see Figure 6), and
then use the toolbox to estimate the intrinsic
(focal length, image center, and pixel skew) and
extrinsic (rotation and translation) parameters
and lens distortion coefficients for each camera.
Because we use a color-based reconstruction
approach, the color responses of the individual
cameras must match. To achieve this, we use a
closed-loop approach that seeks to adjust some
hardware registers in the PGR cameras for opti-
mal color matching and then compensates for
remaining errors in software. By setting the hard-
ware registers to their optimal values during
acquisition, we in effect apply part of the photo-
2D RGB video
Filtering and triangulation
Figure 5. Content creation involves three steps: (a) acquisition of the 2D video of an actual surgical event,
(b) reconstruction of 3D models, and (c) authoring an IEBook.
metric calibration. Residual errors are corrected
during a software refinement step.
To determine optimal values for the eight cam-
eras’ hardware registers, we developed software
that iteratively adjusts the registers while observ-
ing the effects in images of a Gretag Macbeth
“Color-Checker Color Rendition Chart” (see
Figure 6). The software adjusts the parameters in
each camera and color channel such that the
resulting images are as alike as possible.
Specifically, we use Powell’s method for nonlinear
optimization4to minimize a cost function defined
as the weighted sum of the squared differences
between color values, and the variances across the
sampling window of each color sample. The para-
meters of the software refinement step are com-
puted in a onetime final calibration stage.
Synchronized capture. To improve the qual-
ity of the reconstruction results we set the cam-
era cube on top of black fabric as shown in Figure
6. Before capturing an event, we acquire imagery
to estimate the actual color of the fabric, then
during reconstruction we mark as “background”
any scene points that appear to have that color.
To capture an event, we first start the image
capture server software on each of the four Dell
servers, and start a master synchronization pro-
gram that sets off the synchronized capture to
disk. The capture continues for a preset number
of frame sets.
We address server bandwidth and disk space
concerns both by design (the camera-group con-
figuration of Figure 3) and by removing old files
and defragmenting disks prior to capture. Each
640 × 480 camera image is approximately 300
Kbytes. At a capture rate of 15 images per second,
with two cameras per server, we need approxi-
mately 10 Mbytes per second of bandwidth from
the cameras to the server disks, and approxi-
mately 0.5 Gbytes of disk space per minute of
capture per server.
Color space conversion and distortion
removal. As with most single-sensor cameras,
the charge-coupled device (CCD) in the PGR
cameras achieves color imagery using optical fil-
ters that spatially divide the 640 × 480 pixels (pic-
ture elements) among red, green, and blue (RGB).
The CCD in our PGR cameras uses a spatial
arrangement called the Bayer Tile Pattern. As
such, each native captured 8-bit image contains
interspersed data from all three color channels
that a demosaicing algorithm combines to create
a 640 × 480 24-bit RGB image. We developed
software that uses PGR libraries to convert the
native 8-bit Bayer images into 24-bit RGB images.
Our code also uses the Intel Image Processing
Library to remove the lens distortion estimated
during the geometric calibration.
The 3D reconstruction process involves two
reconstruction of 3D points from 2D images
using view-dependent pixel coloring (VDPC)
reconstruction of 3D surfaces from the 3D
points using application-specific point filter-
ing and triangulation to create 3D meshes.
Figure 5b gives an overview of the reconstruction
View-dependent pixel coloring. VDPC is a
hybrid image-based and geometric approach that
estimates the most likely color for every pixel of
an image that would be seen from some desired
viewpoint while simultaneously estimating a
view-dependent 3D model (height field) of the
underlying scene. (Further theory and imple-
mentation details of VDPC and a discussion of
related work are available elsewhere.5) By consid-
ering a variety of factors, including object occlu-
sions, surface geometry and materials, and
lighting effects, VDPC produces results where
other methods fail—that is, in the presence of
textureless regions and specular highlights,
which are common conditions in surgery (skin,
organs, and bodily fluids).
Figure 7 illustrates VDPC’s fundamental con-
cepts. We begin by defining a 3D perspective
voxel grid from the desired viewpoint, which is
typically situated above and looking down into
the camera cube, as Figure 7a demonstrates.
Figure 6. The geometric and
photometric calibration patterns, as
seen from one of the eight cameras.
The patterns shown together here are
actually used separately for
geometric and photometric
As Figure 7b illustrates, for each x-y pixel in
the desired viewpoint image, we effectively tra-
verse the ray away from the desired viewpoint to
estimate the most likely color for that pixel. To
do this, we test each voxel along the ray by back-
projecting it into each of the eight cameras and
looking at the actual camera image color at that
point. We choose one “winner” voxel (the voxel
with the most plausible back-projected appear-
ance in all camera samples). We then use the
median of the winner’s back-projected camera
sample colors as the surface (voxel) color esti-
mate, and the position on the ray as the 3D coor-
dinate of the surface point. We mark all winner
voxels as opaque surfaces, and all others along
the ray between the desired view and the surface
as empty. Like Kutulakos’s and Seitz’s seminal
space-carving work,6this effectively carves away
voxels that do not appear to be part of a surface.
We repeat this volume-sweeping process,
allowing estimated opaque surface points to
occlude other voxels, progressively refining an
estimate of a height field corresponding to the
first visible surfaces. The dashed box in Figure 5b
illustrates this cycle.
Two aspects of VDPC contribute to its relative
robustness in the presence of textureless and
First, we use a view-dependent smoothness con-
straint, assuming that 3D surfaces are more likely
to be continuous without abrupt
change in depth. We implement the
constraint with a new volumetric
formulation of the relatively well-
known disparity gradient principle.
VDPC uses this constraint, updating
the likelihoods as it progressively
refines the surface voxel model.
Second, VDPC uses a physically
based consistency measure. When
sweeping along x-y desired view-
point rays, we declare the voxel
with the most plausible back-pro-
jected appearance the winner. One
measure of plausibility is the vari-
ance of the back-projected color
samples—the smaller the variance
the more likely a voxel is to be a sur-
face voxel. Our physically based
consistency measure goes farther,
allowing for back-projected sample
color distributions that are plausible
in the presence of specular high-
lights on many types of surfaces. In
these cases, the colors of back-projected samples
occurring near specular reflections of a light
source tend toward the light’s color. In fact, in
the presence of these specular highlights, the
eight back-projected sample colors will lie along
a line in the RGB color space, extending from the
surface’s inherent color toward the light’s color.
By allowing for this specular distribution of back-
projected sample colors, we can detect surfaces
where simple variance tests would fail.
When we apply VDPC to a frame set, we get a
static surface voxel model from the desired view-
point’s perspectivethat is, a height field. In fact,
because some x-y rays finish without winner vox-
els and some are declared part of the background,
the result is a sparse surface voxel model. We use
the estimated camera geometry to transform this
into a 3D point cloud in the camera cube space,
and repeatedly apply the overall approach to all of
the frame sets to obtain a dynamic 3D point cloud.
Filtering and triangulation. After recon-
struction, some holes will exist in the 3D point
data, and some 3D points will visually appear to
be outliers. If we filter and triangulate the 3D
points where we believe there are indeed sur-
faces, we can obtain more continuous dynamic
models, further improving the appearance.
To this end, we developed new postprocessing
methods to filter and triangulate the dynamic
to pixel (x, y) in the
desired view image
Figure 7. View-dependent pixel coloring (VDPC): (a) the desired viewpoint and perspective
voxel grid concepts, and (b) the color consistency check for a particular voxel along a ray
from the desired viewpoint image. The arrows extending from the hand toward the cameras
depict the reflectance or imaging of the surface (voxel) into or by the cameras.
sparse 3D point data. Figure 8 shows some results.
Although we could use various off-the-shelf
methods, VDPC allows for more specific con-
straints. For example, because the reconstructed
points are on a 2D view-dependent grid, we can
use 2D algorithms for filtering and triangulation.
In addition, VDPC assigns a unique color to
points estimated to be on a surface’s edges, letting
us “break” meshes at edges and helping us iden-
tify holes within the reconstructed shapes.
Our postprocessing filtering and triangulation
software applies three steps in order:
1.Median filtering. The software uses a 5 × 5
median filter to smooth the z values. For each
sample, the software inspects the 24 grid loca-
tions in the surrounding 5 × 5 x-y square. If at
least three of these locations contain samples
(including edge points), we replace the origi-
nal sample’s z value with the median value.
2.Hole filling. Exploiting the edge-color contour-
ing, the software fills holes by examining
neighboring samples. For each unpopulated
sample located in a hole, the software uses a 3
× 3 box filter to average both color and depth
values from up to eight neighboring samples.
3.Triangulation. Finally, the software examines
the array in square groups of 2 × 2 samples. If
any group is fully populated, we can use two
possible triangulation techniques. We pick
the one that minimizes the color difference
along the diagonal.
Our current approach to authoring includes a
novel combination of 2D and 3D interaction
techniques. The primary motivation for a hybrid
2D/3D approach is to provide a familiar and tan-
gible means of sketching (the notepad paradigm)
while simultaneously offering a natural and
immersive means of viewing the dynamic 3D
data and the evolving IEBook.
Figure 5c illustrates the authoring process.
Figure 9a shows an author annotating an IEBook
in the VR-Cube, and Figure 9b is a screen shot of
the authoring interface on the Tablet PC.
Using VCR-like time controls (see the right
Figure 8. Filtering and
(b) triangulation with
filtering, (c) filtered
points colored in green
and filled holes in red,
and (d) absolute filter
correction values as a
height field. The highest
spikes correspond to
pointy outliers in (a).
Figure 9. Authoring an
IEBook: (a) an author in
the VR-Cube and (b) a
screen shot of the inter-
face on the Tablet PC.
side of Figure 9b), the author navigates time in
the captured sequence, looking for an interesting
or important event. The author moves to a view-
point where he or she has a good view of the sur-
geon’s actions at that moment, and takes a
snapshot of it using a button on the Tablet PC.
Using the same Tablet PC interface, the author
can highlight features, annotate the snapshot,
and save the results to a virtual gallery in the
IEBook. The author can arrange the snapshots
hierarchically by dragging their titles on the
Tablet PC application.
Unfortunately, conventional methods for
selection and highlighting in 2D images do not
accommodate depth or time. We therefore adapt-
ed some conventional methods to enable the
selection and highlighting of dynamic 3D
point/mesh data.7Specifically, we implemented
three distinct dynamic 3D highlighting para-
digms: marquee, freeform, and fill, which Figure
10 illustrates. The implementations preserve high-
lights throughout a sequence and are adaptive,
reacting to changes in the 3D topology over time.
To create a marquee highlight, the author
drags a rectangle (or oval) around an area of
interest in a 2D snapshot. The authoring appli-
cation renders the selection into the current
frame of the dynamic 3D data, darkening all
points outside the selected area by an adjustable
amount. The effect is akin to a square (or oval)
spotlight illuminating the region of interest. The
author can make as many selections as desired in
the current frame. When the user presses play,
the authoring software attempts to highlight the
same regions in subsequent frames until the ani-
mation pauses. We also implemented a surface-
specific variation that only highlights actual
surfaces within the marquee region, omitting
holes or background regions.
The freeform highlight is a paintbrush-style
tool for creating colored marks on a fixed surface
within the point cloud. To create a freeform
highlight, the author clicks and drags the mouse
to create a mark in the tablet display. When the
author releases the mouse, the system renders the
highlight onto the current frame as a tint of the
selected color. When the author highlights sub-
sequent frames, portions of the highlight are
obscured whenever their corresponding surface
points are obscured. The freeform highlight is
surface specific in that the highlight’s location is
based entirely on the surface points’ positions.
The fill highlight is similar to a “paint bucket”
tool. It applies a tint of the current highlight
color to a group of contiguous points around a
selected location, all of which have similar depth
and color values. Like the freeform highlight, the
fill is surface specific. Unlike the freeform high-
light, however, the fill highlight can shrink and
stretch, letting it compensate for small amounts
of movement in the highlighted region.
The IEBook stores the highlight information
as a series of 2D Boolean matrices indicating
which vertices should be highlighted in each
frame. Currently, we have 10 independent high-
light layers, and can assign each to a different
color. We can also hide the layers, letting the user
see the results with or without their effects.
Our primary paradigm for experiencing an
IEBook is head-tracked stereo 3D, in which a stu-
dent uses the VR-Cube and a subset of the author-
ing interface. This IEBook environment combines
immersive 3D imagery with Tablet PC-based nav-
igation and a hierarchy of annotated snapshots.
A student views the snapshot gallery in the
VR-Cube, holding the Tablet PC as in Figure 9.
Figure 10. Examples of our dynamic 3D highlighting paradigms: (a) marquee, (b) freeform, and (c) fill. The tip of a freeform arrow in
(b) is partially occluded by the surgeon’s hands, illustrating the 3D nature of the highlights. Note that this data set has not been
filtered and triangulated.
The user can navigate to a specific point in time
using the VCR-like controls or by choosing a
snapshot in the hierarchy. A blue frame provides
feedback about the active snapshot during the
selection process. When the user selects a snap-
shot, the model jumps to the time associated
We built the software for our hybrid display
system prototype (see the “Visualizing Procedures
Up Close and All Around” sidebar) on the same
scene graph software used in the Barco TAN VR-
Cube implementation. However, we added a ren-
dering mode to allow rendering in both a
head-mounted display and with projectors illu-
minating disjoint and casually arranged (not in a
cube) planar surfaces. (We currently use this
equipment to show only the dynamic 3D
points/meshes from the reconstruction phase of
content creation and not the results of authoring.)
Although we believe the head-tracked stereo
3D systems and VR-Cube provide the greatest
sense of immersion, they are typically available
only in visualization labs. So that more people
can experience our system, we made it available
on the Web. In consultation with our medical
partners, we chose three media: anaglyphic
stereo images, anaglyphic movies, and nonstereo
but dynamic (in space and time) VRML models.
We developed software to generate anaglyphic
stills and movies from the 3D point/mesh data
and separate software that creates “boxed” VRML
data sets. Each of the latter features a 3D VCR-like
user interface to play and step through a
sequence, including a slider bar for random
access similar to typical computer-based movie
Figure 11 shows several views of a dynamic
reconstruction of Bruce Cairns performing a
mock procedure to manage blunt liver trauma
(from Figure 2b), and another coauthor manipu-
lating Rhesus monkey and human skulls.
Figure 12 (next page) shows a preliminary
IEBook in the VR-Cube. The six annotated snap-
shots correspond to six (mock) items of interest.
The third item is selected and playing. Zoom
boxes magnify the imagery in the screen capture
Figure 11. Reconstruction results: (a) a mock procedure to manage blunt liver trauma, (b) manipulating a Rhesus monkey skull, and
(c) a human skull from a 12-year-old child. The human skull is a teaching artifact that can be separated into multiple parts. Both
skull reconstruction sequences include simulated dynamic shadows.
for several items. (The annotations are clear to an
actual user, but do not reproduce well in a rela-
tively small image.)
One of the skills most needed by new sur-
geons, yet difficult to learn, is suturing. Knot-
tying skills in particular can be critical—the
wrong knot can cause a wound to open or result
in tissue damage. The surgeons we know agree
that 2D knot-tying images and knot-tying
movies are woefully inadequate. We therefore
attempted to acquire and reconstruct some fun-
damental knot-tying demonstrations. We can
reconstruct hands and moderately thick rope,
and have done so for a few basic surgical knots as
well as some common sailing knots. Figure 13
shows an anaglyphic stereo movie and the VRML
interface of a ring knot tying reconstruction.
Results showing basic sailing knots
and common surgical knots are
available at http://www.cs.unc.
reconstructions/. We will continue
to add reconstruction results as we
We have shown several surgeons
our lab-based (stereo and head-
tracked) and Web-based (movies
and VRML) reconstructions (such as
those in Figure 11, not the complete
IEBook results in Figure 12). The
surgeons’ reactions have been very
Most of the surgeons cited size,
resolution, and visibility of the recon-
structions as the primary limitations,
which is no surprise to us at the pro-
ject’s relatively early stage. We did not optimize the
choice and arrangement of cameras in our camera
cube (see Figure 2) for any particular procedure.
Rather, we based our choices on what was practical
for a proof-of-concept system supporting recon-
structions of modest detail throughout a modest
As the “Visualizing Procedures Up Close and
All Around” sidebar explains, both the procedure
and the surrounding events are important.
Achieving appropriate resolution in terms of
reconstructions is a huge challenge. Ideally, we
would have 3D reconstructions on the order of
millimeters over the entire operating room,
which might be 5–10 meters per side (125–1,000
cubic meters). Because this is currently impracti-
cal, near-term future challenges will include
Figure 12. Screen shot
from the VR-Cube
arranged in an IEBook
virtual gallery. Zoom
boxes magnify three of
the snapshots, and the
blue frame and red
marker indicate the
Figure 13. Ring knot
tying: (a) anaglyphic
stereo movie (“red over
right”), and (b) VRML
interface. The movie
dynamic shadows; the
VRML model does not.
choosing camera configurations and algorithms
that can achieve high-resolution reconstructions
where most critical, and relatively low-resolution
reconstructions everywhere else.
Despite our achievements over several years,
we are still in the early stages of a long-term
effort. Many problems remain to be solved, and
there is much work to be done to achieve the
scale, fidelity, flexibility, and completeness that
In the near term, we will continue to recon-
struct and process various objects and mock pro-
cedures. For example, we are procuring some
rapid-prototype models of a real skull for a mock
sagittal synostosis procedure. Soon we hope to
provide anaglyphic stereo 3D data sets on the
Web, letting viewers control the viewpoint and
time. We are trying to solve problems with the
network connection to the Tablet PC used for
authoring and viewing. Although a wireless par-
adigm is clearly desirable, the VR-Cube hardware
seems to interfere with the wireless connection.
A long-term primary goal is to scale our system
up to enable the acquisition and reconstruction
of (at least) a surgical table, the surrounding area,
and the involved medical personnel. In fact, with-
in several years we plan to equip an intensive care
unit with numerous heterogeneous cameras to
capture a variety of medical procedures. This
involves solving nontrivial problems such as res-
olution, visibility, computational complexity, and
massive model management. To assist with cam-
era placement, we are working on a mathemati-
cal and graphical tool to help estimate and
visualize acquisition information/uncertainty
throughout the acquisition volume for a particu-
lar candidate set of cameras.
In light of our goals for fidelity and scale, we
continue to improve our existing methods for 3D
reconstruction and to investigate new methods.
For example, we might combine pre-acquired
laser scans of an operating room with camera-
based dynamic reconstructions so we can better
allocate cameras for dynamic events. We are also
investigating possibilities for capturing real-time
data from medical monitoring equipment, and
plan to include it as metadata in an IEBook.
Whereas in the past we chose to hold off on
audio acquisition, instead attempting to address
the more difficult problems related to visual
reconstruction, we plan to acquire audio as well.
We are also rethinking our choice of a Tablet PC
as the primary interface to the immersive author-
ing system. The tablet display is hard to read while
wearing stereo glasses, and both hands are typical-
ly busy holding the tablet and the stylus, so remov-
ing the stereo glasses when necessary is not an
attractive solution. In our current paradigm,
authors choose a view with their heads (looking at
the data of interest) while trying to use the tablet
stylus to initiate snapshots, which is awkward. One
possibility is to track the Tablet PC so users can
choose snapshots in a “viewfinder” mode.8
Finally, we hope to increase the impact on the
medical community by making complete IEBooks
available on the Web. The primary difficulty here
is in determining which interaction techniques
are appropriate and how to implement them.
Rather than simply “dumbing down” the fully
immersive interfaces, we want to use the best
interfaces for each paradigm and authoring tools
that appropriately target each.
At the University of North Carolina, we
acknowledge Marc Pollefeys for VDPC collabora-
tion; Jim Mahaney and John Thomas for techni-
cal support; and surgeons Ramon Ruiz and
Anthony Meyer for general collaboration. At
Brown University, we thank Melih Betim and
Mark Oribello for their systems and video support.
This research was primarily supported by US
National Science Foundation Information
Technology Research grant IIS0121657, and in
part by US National Library of Medicine contract
N01LM33514 and NSF Research Infrastructure
1. H. Fuchs et al.,“Virtual Space Teleconferencing
Using a Sea of Cameras,” Proc. 1st Int’l Symp.
Medical Robotics and Computer Assisted Surgery,
Shadyside Hospital, Pittsburgh, 1994, pp. 161-167.
2. A. van Dam et al., “Immersive Electronic Books for
Teaching Surgical Procedures,” Telecomm.,
Teleimmersion, and Telexistence, S. Tachi, ed.,
Ohmsha, 2002, pp. 99-132.
3. J.Y. Bouguet, “Camera Calibration Toolbox for
4. W. Press et al., Numerical Recipes in C: The Art of
Scientific Computing, 2nd ed., Cambridge Univ.
5. R. Yang, View-Dependent Pixel Coloring—A Physically
Based Approach for 2D View Synthesis, doctoral disser-
tation, Univ. of North Carolina at Chapel Hill, 2003.
6. K. Kutulakos and S. Seitz, “A Theory of Shape by
Space Carving,” Int’l J. Computer Vision, vol. 38, no.
3, 2000, pp. 199-218.
7. D.M. Russo, “Real-Time Highlighting Techniques
for Point Cloud Animations,” master’s thesis, Brown
8. M. Tsang et al., “Boom Chameleon: Simultaneous
Capture of 3D Viewpoint, Voice, and Gesture
Annotations on a Spatially Aware Display,” Proc.
15th Ann. ACM Symp. User Interface Software and
Technology, ACM Press, 2002, pp. 111-120.
Greg Welch is a research associate
professor of computer science at
the University of North Carolina
(UNC) at Chapel Hill. His research
interests include tracking and
sensing systems for virtual envi-
ronments and 3D telepresence. Welch has a PhD in
computer science from UNC-Chapel Hill. He is a mem-
ber of the IEEE Computer Society and the ACM.
Andrei State is a senior research
scientist in the Department of
Computer Science at UNC-Chapel
Hill. His technical interests
include 3D graphics and mixed
reality. State has a Dipl.-Ing.
degree from the University of Stuttgart and a master’s
degree in computer science from UNC-Chapel Hill.
Adrian Ilie is a graduate research
assistant in the Department of
Computer Science at UNC-
Chapel Hill. His research interests
include camera placement for 3D
reconstruction, photometric cali-
bration, and image processing. Ilie has an MS in com-
puter science from UNC-Chapel Hill.
Kok-Lim Low is a doctoral can-
didate in the Department of
Computer Science at UNC-
Chapel Hill. His research interests
include projector-based render-
ing, augmented reality, and 3D
imaging and modeling using active range sensing. Low
has an MS in computer science from UNC-Chapel Hill.
Anselmo Lastra is an associate
professor of computer science at
UNC-Chapel Hill. His research
interests are computer graphics,
specifically image-based model-
ing and rendering, and graphics
hardware architectures. Lastra has a PhD in computer
science from Duke University.
Bruce Cairns is an assistant pro-
fessor at UNC-Chapel Hill and
director of the North Carolina
Jaycee Burn Center. His research
interests include trauma, burns,
critical care, cellular immunolo-
gy, and shock. Cairns has an MD from the University
Herman Towles is a senior
Department of Computer Science
at UNC-Chapel Hill. His research
interests include graphics hard-
ware architecture, large-format
projective displays, 3D scene reconstruction, and tele-
presence. Towles has an MS in electrical engineering
from the University of Memphis.
Henry Fuchs is the Federico Gil
Professor of Computer Science,
adjunct professor of biomedical
engineering, and adjunct profes-
sor of radiation oncology at
UNC-Chapel Hill. His research
interests include computer graphics and vision and
their application to augmented reality, telepresence,
medicine, and the office of the future. Fuchs has a PhD
in computer science from the University of Utah. He is
a member of the National Academy of Engineering and
a Fellow of the ACM and of the American Academy of
Arts and Sciences.
Ruigang Yang is an assistant pro-
fessor in the Computer Science
Department at the University of
Kentucky. His research interests
include computer graphics, com-
puter vision, and multimedia.
Yang has a PhD in computer science from UNC-Chapel Download full-text
Hill. He is a recipient of the NSF Career Award, and is a
member of the IEEE and the ACM.
Sascha Becker is a research sci-
entist for the Brown Computer
Graphics Group and a senior soft-
ware engineer at Laszlo Systems
in San Mateo, California. Her
work focuses on using technolo-
gy as a tool for art, education, and play. Becker has a BA
in computer science from Brown University. More
information is available at http://sbshine.net.
Daniel Russo is a software devel-
oper at Filangy, a start-up intent
on revolutionizing Web search-
ing. Russo has an ScM in comput-
er science from Brown University.
Jesse Funaro recently earned an
ScM in computer science from
Brown University, with research
in optimization techniques for
search result ranking.
Andries van Dam is the vice
president for research and profes-
sor of computer science at Brown
University. His principal research
interests are computer graphics,
computer-based education, and
hypertext and electronic books. Van Dam has a PhD in
electrical engineering from the University of
Pennsylvania. He is a member of the National Academy
of Engineering and a Fellow of the ACM and the IEEE.
Readers may contact Greg Welch at the University of
North Carolina at Chapel Hill, Department of
Computer Science, CB# 3175, Sitterson Hall, Chapel
Hill, NC 27510-3175; firstname.lastname@example.org.
CERTIFIED SOFTWARE DEVELOPMENT PROFESSIONAL PROGRAM
Apply now for the 1 April—30 June test window.
Visit the CSDP web site at www.computer.org/certification
or contact email@example.com
Doing Software Right
Demonstrate your level of ability in relation to your peers
Measure your professional knowledge and competence
Certification through the CSDP Program differentiates between you and other software
developers. Although the field offers many kinds of credentials, the CSDP is the only one
developed in close collaboration with software engineering professionals.
“The exam is valuable to me for two reasons:
One, it validates my knowledge in various areas of expertise within the software field, without regard to specific
knowledge of tools or commercial products...
Two, my participation, along with others, in the exam and in continuing education sends a message that software
development is a professional pursuit requiring advanced education and/or experience, and all the other
requirements the IEEE Computer Society has established. I also believe in living by the Software Engineering
code of ethics endorsed by the Computer Society. All of this will help to improve the overall quality of the
products and services we provide to our customers...”
— Karen Thurston, Base Two Solutions