Permission to make digital or hard copies of part or all of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed for
profit or direct commercial advantage and that copies show this notice on the first page or
initial screen of a display along with the full citation. Copyrights for components of this
work owned by others than ACM must be honored. Abstracting with credit is permitted. To
copy otherwise, to republish, to post on servers, to redistribute to lists, or to use any
component of this work in other works requires prior specific permission and/or a fee.
Permissions may be requested from Publications Dept., ACM, Inc., 1515 Broadway, New
York, NY 10036 USA, fax +1 (212) 869-0481, or firstname.lastname@example.org.
High Dynamic Range Display Systems
, Wolfgang Heidrich
, Wolfgang Stuerzlinger
, Greg Ward
, Lorne Whitehead
, Abhijeet Ghosh
, Andrejs Vorozcovs
The University of British Columbia,
Figure 1: From left to right: our projector-based display showing an HDR image; our LED-based HDR display showing a discrete and a
smooth intensity ramp (the top half of the discrete ramp and the bottom half of the smooth ramp have each been covered by a 1% transparent
ﬁlter to illustrate high luminance content on the left side of the image, which cannot be captured by the camera); a color-coded original HDR
image; HDR photograph taken off the screen of our projector-based system; HDR photograph taken off a conventional monitor displaying
the tone-mapped image.
The dynamic range of many real-world environments exceeds the
capabilities of current display technology by several orders of mag-
nitude. In this paper we discuss the design of two different display
systems that are capable of displaying images with a dynamic range
much more similar to that encountered in the real world. The ﬁrst
display system is based on a combination of an LCD panel and
a DLP projector, and can be built from off-the-shelf components.
While this design is feasible in a lab setting, the second display
system, which relies on a custom-built LED panel instead of the
projector, is more suitable for usual ofﬁce workspaces and commer-
cial applications. We describe the design of both systems as well as
the software issues that arise. We also discuss the advantages and
disadvantages of the two designs and potential applications for both
Keywords: Hardware – Novel Display Technologies; Rendering
– Perceptually Based Rendering; Image and Video Processing –
Image Processing; Methods and Application – Signal Processing;
Hardware – Framebuffer Algorithms.
CR Categories: B.4.2 [INPUT/OUTPUT AND DATA COMMU-
NICATIONS]: Input/Output Devices—Image display; I.3.3 [COM-
PUTER GRAPHICS]: Picture/Image Generation—Display algo-
rithms; I.3.4 [COMPUTER GRAPHICS]: Graphics Utilities—
Device drivers; I.4.0 [IMAGE PROCESSING AND COMPUTER
VISION]: General—Image displays.
In the past few years, the limited dynamic range of both imaging
devices and displays has received a lot of attention in the computer
graphics community. Algorithms have been developed for captur-
ing both photographs [Mann and Picard 1994; Debevec and Ma-
lik 1997; Robertson et al. 1999; Mitsunaga and Nayar 1999] and
videos [Kang et al. 2003] with extended dynamic range.
Simultaneously, tone mapping operators have also been devel-
oped for compressing the dynamic range so that the images can be
displayed on the familiar 8 bit/channel displays with typical con-
trast ratios of about 300 : 1, including any conventional Cathode
Ray Tube (CRT), Liquid Crystal (LCD), and projector-based dis-
play. While these tone mapping operators (e.g. [Schlick 1994; Lar-
son et al. 1997; Tumblin and Turk 1999; Durand and Dorsey 2002]
among others) allow for displaying high-dynamic-range (HDR) im-
ages in a recognizable and even aesthetically pleasing way, nobody
would confuse a photograph rendered in this fashion with, say,
watching the same scene through a window. The dynamic range
of conventional displays is simply insufﬁcient to create the optical
sensation of watching a real sunset or driving a car into oncoming
trafﬁc at night. Note that this is not just an issue of top intensity:
simply increasing the brightness of a conventional display would
wash out the dark tones and turn them into a medium gray. What is
needed is a signiﬁcant expansion of the contrast or dynamic range
of the display.
In this paper we describe two alternative designs for HDR dis-
play systems. We have built prototypes of both, and discuss both
the optical design, and related software issues such as display cali-
bration and the rendering of HDR images on both displays.
Both display systems are based on the fundamental idea of using
an LCD panel as an optical ﬁlter of programmable transparency to
modulate a high intensity but low resolution image from a second
display. For example, assume we have any display with a contrast
range of c
: 1 between the darkest and the brightest intensity pro-
© 2004 ACM 0730-0301/04/0800-0760 $5.00
ducible by that display. If we now put an LCD panel with a contrast
ratio of c
: 1 in front of the ﬁrst one, then the (theoretical) contrast
of the combined system is (c
) : 1. In practice, the ﬁrst display
needs to be able to produce a very high intensity image, because
color LCD panels only have a transparency of about 3-8%, even
when switched to ‘white’, so that most energy is actually absorbed.
Another reason for using a display with a very high base intensity
is that a lot of the HDR images we would like to show have, in fact,
very bright regions in them.
Based on this principle, we have derived two alternative designs
for HDR displays. In the ﬁrst design (Section 4), a video projector
based on Digital Light Projector (DLP) technology serves as the
base display. In this version, we directly focus the projector onto
the back of the LCD panel. Since the illuminated area is much
smaller than during regular use of a projector, the light density is
dramatically improved, yielding the high top intensities that we are
aiming for. While this design works well in a laboratory setting,
it has several drawbacks that restrict its use for a wider class of
applications. In particular, these are a large form factor, signiﬁcant
power consumption and heat development, as well as calibration
To overcome these issues, we have devised a second design (Sec-
tion 5), in which the projector is replaced with a low-resolution
array of ultra-bright LEDs. The intensity of every LED can be pro-
grammed individually, yielding a low resolution version of the de-
sired image. High frequency features are introduced by attaching
a high-resolution LCD panel to the front of this LED array, and
adjusting its transparency accordingly. This design makes use of
results from psychophysics, which show that very high contrast, al-
though important on a global scale, cannot be perceived by humans
at high spatial frequencies (see Section 3).
The two displays we developed have dynamic ranges well be-
yond 50, 000 : 1, and a maximum intensity of 2700cd/m
, respectively. This compares to a typical dynamic
range of about 300 : 1, and a maximum intensity of about
for a typical desktop display.
The design of both systems and their appropriate rendering algo-
rithms, as well as the advantages and disadvantages are detailed in
Sections 4 and 5. In Section 6 we discuss possible applications for
our display technology before we conclude with some remarks and
future research directions in Section 7.
2 Related Work
The class of image processing techniques for coping with the dis-
crepancy between real world luminances and those that ﬁt within
the limited gamut of a conventional output device is collectively
called tone-mapping. Tumblin and Rushmeier  introduced
this concept to computer graphics, though their early work did not
address dynamic range limitations per se. The ﬁrst tone-mapping
operator to tackle dynamic range reduction was Chiu et al. ,
who used a spatially varying exposure ramp over the image. How-
ever, this approach led to disturbing “reverse gradients” typically
seen as halos around light sources. Later work by Larson et
al.  returned to a global operator for dynamic range reduction
based on histogram adjustment to avoid these artifacts, with local
variations to simulate disability glare due to high contrast bound-
aries in a scene. Pattanaik et al.  developed what some re-
searchers consider the ultimate still image operator based on the
human visual system, incorporating color adaptation, local con-
trast, and dynamic range. However, even this operator exhibited
some reverse-gradient effects near high contrast boundaries due to
its local spatial ﬁlters, leading other researchers to take a different
The basic challenge for a spatially varying tone-mapping opera-
tor is that it needs to reduce the global contrast of an image without
affecting the local contrast to which the human visual system (HVS)
is sensitive. To accomplish this, an operator must segment the HDR
image, either explicitly or implicitly, into regions the HVS does not
correlate during dynamic range reduction. The ﬁrst researchers to
successfully accomplish this in an automatic tone-mapping were
Tumblin and Turk  with their LCIS operator. However, LCIS
sometimes produces odd-looking images, which bear little rela-
tion or resemblance to the original scene brightnesses. More re-
cent operators by Ashikhmin , Fattal et al. , Reinhard
et al. , Durand & Dorsey , and Choudhury & Tum-
blin , are much more successful in separating contrast differ-
ences that matter to vision from those that do not.
Regrettably, none of these techniques has been validated or sup-
ported by psychophysical research, so the resulting images remain
as the only evidence that these methods have any real validity for re-
producing our experience of an HDR scene on a low dynamic range
device. Most of the methods contain free parameters that must be
set by the user based on preference. Usually, little can be said about
the visual impact that these parameters have on the image in terms
of visibility, contrast, brightness, or human visual response in gen-
HDR displays offer two beneﬁts in this area. The ﬁrst, immedi-
ate beneﬁt is to researchers, who until now have had no means for
the controlled display of dynamic HDR imagery in their studies.
Our HDR displays are already helping researchers to test out their
hypotheses regarding the effects of tone-mapping on HDR scenes.
The longer term beneﬁt will be felt when HDR displays penetrate
the professional and eventually the consumer markets, reducing or
eliminating the need for tone-mapping. Although we expect tone-
mapping to continue to be a requirement for many types of out-
put, such as hard copy, the availability of HDR displays will likely
reduce the need for dynamic range compression for many critical
applications in the years to come.
3 Remarks on Human Perception
The human visual system has tremendous capabilities but also some
limitations. Some of these limitations are an integral part of the
theory underlying the HDR devices presented in this paper and the
following sections will describe these in detail. In general, our eye
has evolved to deal with the vast dynamic range available to us in
our daily environment, ranging from starlight to sunlight over at
least an eight order of magnitude luminance range. To cope with
this range, the eye uses a complex adaptation system. For the pur-
pose of this paper, we use a simple model of adaptation with two
time scales: those mechanisms working at time scale of the order of
minutes and those with a shorter time scale. The former are of little
interest from the point of view of HDR display development. The
latter are very interesting as they are the primary reason why current
displays cannot provide realistic representations of real world HDR
scenes. The eye can capture approximately 5 orders of magnitude
of dynamic range effectively simultaneously. No conventional dis-
play technology comes close to this. Yet, there are limitations to
this capability as described below.
3.1 Local Contrast Perception
While we can see a vast dynamic range across a scene, we are un-
able to see more than a small portion of it in small regions (cor-
responding to small angles). Different researchers report differ-
ent values for the threshold past which we cannot make out high
contrast boundaries, but most agree that the maximum perceivable
contrast is somewhere around 150 : 1 [Vos 1984]. Scene contrast
boundaries above this threshold appear blurry and indistinct, and
the eye is unable to judge the relative magnitudes of the adjacent
Angle (in degrees)
Point Spread Function of the Human Eye
Figure 2: The point spread function of the human eye according to
This inherent limitation can be explained locally by the scatter-
ing properties of the eye. From Moon & Spencer’s original work
on glare , we know that any high contrast boundary will scat-
ter at least 4% of its energy on the retina to the darker side of the
boundary, obscuring the visibility of the edge and details within a
few degrees of it (Figure 2). If the contrast of an edge is 25 : 1, then
details on the darker side will be competing with an equal amount
of light scattered from the brighter side, reducing visible contrast by
a factor of 2 in the darker region. When the edge contrast reaches a
value of 150 : 1, the visible contrast on the dark side is reduced by
a factor of 12, rendering details indistinct or invisible.
However, we cannot claim high contrast content has no effect –
clearly it does. An observer will notice when one region is much
brighter than another, both by the challenge it creates in viewing
the boundary, and by the accommodation that goes on when shift-
ing from side to side. When the threshold is very large, observers
may even experience discomfort as they attempt to see detail near
a bright source, as any driver knows from their nighttime travels.
A photographic print of oncoming headlights is merely an allusion
to the real experience – it cannot duplicate the visceral experience
of glare, or reproduce the effect it has on a human observer. It is
exactly this kind of experience that an HDR display can uniquely
The HDR display technology described in this paper only ex-
ploits the inability of humans to see detail in the immediate vicinity
of a high-contrast boundary; it makes no assumptions about our
overall response to varying brightnesses. Relative (and even ab-
solute) brightnesses are maintained, and edges will be reproduced
exactly when they are below the maximum contrast of the front dis-
play – about 400 : 1 in our current prototype. Only when this range
is exceeded is some ﬁdelity lost near high contrast boundaries, but
this effect is well below the detectable threshold, and has not been
visible in any of our experiments [Seetzen et al. 2003].
3.2 Just Noticeable Difference Steps vs. Contrast
and Bit Depth
A key question to ask in designing any display system is: how many
distinct input/output levels are necessary to cover the desired range
without banding or similar quantization artifacts? For conventional
displays, this question is often answered by considering a single
viewer adaptation level and the number of bits required to represent
suitable steps on a particular gamma curve. This may be adequate
if the dynamic range being considered is small, but fails when a
display is capable of levels much brighter and much darker than
To answer this question in the context of a HDR displays, we
turn to psychophysical research in the area of Just Noticeable Dif-
ferences. A JND is the smallest detectable luminance difference at
a given luminance level. Visual psychologists have established a
complete theory for calculating JNDs over the luminance range rel-
evant for our HDR displays [Barten 1992; Barten 1993], which al-
lows us to establish the useful luminance level of the HDR display.
Adding a JND to a particular luminance level effectively deﬁnes
the next useful step on the luminance scale of the display since it is
clearly redundant to provide addressable luminance levels between
those two levels if the eye cannot perceive any difference.
Luminance in cd/m
# Just Noticeable Difference Levels for Different Maximum Luminance Levels
Figure 3: The number of just noticeable difference (JND) steps
for different maximum intensities according to the model by
Based on Barten’s original work, an analytical formula for JNDs
was derived for a DICOM standard  (see Figure 3). This
model predicts 962 JND steps in the luminance range of the pro-
jection based HDR display described in Section 4 and 1139 JND
in the luminance range of the LED based HDR display (Section 5).
Our goal therefore is to reproduce at least this many steps on each
display from the darkest to brightest output level, and in both cases
our step sizes are maintained well below a JND throughout the lu-
minance range. In the following sections we will use the notion of
JNDs instead of contrast and bit depth to characterize the perfor-
mance of the HDR systems, but continue to use these more familiar
terms for the individual components (LCD and projector) that make
up the HDR display.
Recently, Muka and Reiker  have argued that, for conven-
tional displays with a typical dynamic range of 300 : 1 or so, an
8-bit representation of images is sufﬁcient for medical diagnosis.
They argue that the difference between an 8-bit digital display and
a 10-bit or higher bit depth is minimal, and perhaps not noticeable
at all. However, as the range of displayable luminances increases,
so does the number of JND steps required to cover that range, which
is reﬂected in the numbers presented above.
4 System 1: Projector-based Display
As outlined in the introduction, the ﬁrst HDR display system we
built modulates the image from a projector with a transmissive
LCD. This system is detailed in the following.
4.1 Hardware Setup
In a conventional LCD, two polarizers and a liquid crystal are used
to modulate the light coming from a uniform backlight, typically a
ﬂuorescent tube assembly. The light is polarized by the ﬁrst polar-
izer and transmitted through the liquid crystal where the polariza-
tion of the light is rotated in accordance with the control voltages
applied to each pixel of liquid crystal. Finally, the light exits the
LCD by transmission through the second polarizer. The luminance
level of the light emitted at each pixel is controlled by the polariza-
tion state of the liquid crystal. It is important to point out that LCDs
cannot completely prevent light transmission - even at the darkest
state of a pixel, light is emitted and as such the dynamic range of an
LCD is deﬁned by the ratio between the light emitted at the bright-
est state and the light emitted in the darkest state. For a high end
LCD, this ratio is usually around 300 : 1, with monochromatic spe-
cialty LCDs (e.g. those for medical imaging) going up to 700 : 1.
The luminance range of the display can easily be adjusted by con-
trolling the brightness of the backlight, but the dynamic range ratio
will remain the limiting factor. In order to maintain a reasonable
‘black’ level of about 1cd/m
, the LCD is thus limited to a maxi-
mum brightness of about 300cd/m
The basic modiﬁcation introduced by the HDR technology in-
volves inserting a second light modulator and increasing the bright-
ness of the backlight. These two modulators in series provide an ex-
tremely dark state with a very low light emission, which then makes
it possible to increase the brightness of the backlight dramatically
without losing the ‘black’ state. Optically, this series of modulators
results in multiplication of the individual dynamic ranges.
For the projector-based HDR display presented in this paper, the
backlight and the ﬁrst modulator are combined into a single DLP
using a Digital Mirror Device with a dynamic range of 800 : 1.
The three central components of the HDR display are then the pro-
jector, the LCD and the optics that couple the two. Using these
components, each image on the HDR display is the result of modu-
lated light coming from the projector which is directed onto the rear
of the transmissive LCD by the optics system, modulated a second
time by the LCD, and properly diffused for viewing.
The projector used in the HDR display is an Optoma DLP
EzPro737 digital mirror projector. To reduce unnecessary light loss,
we have removed the color wheel from the projector, resulting in
a monochrome display system with a threefold increase in bright-
ness due to the absence of the color ﬁlters. New control electronics
have been integrated into the commercially available projector to
re-synchronize it in absence of this color wheel.
The LCD panel is a 15” XGA color LCD made by Sharp (Sharp
LQ150X1DG0). It is driven by an EarthVision AD2 LCD con-
troller, which allows a direct VGA connection. The LCD panel has
been separated from the conventional backlight and all of the opti-
cal layers behind the display have been removed to create a trans-
missive image modulator.
The optics used in the HDR display include the conventional
projection lens of the EzPro projector, and a Fresnel lens directly
behind the LCD display to collimate the projected light into a nar-
row viewing angle for maximum brightness of the HDR display and
to avoid color distortion due to diverging light passing through the
color ﬁlters of the LCD. Finally, a standard LCD diffuser has been
used to redistribute the collimated light into a reasonable viewing
All three components have been installed in a single housing
with appropriate alignment mechanisms to create a close match-
ing of the DLP and LCD pixels. The alignment can be ﬁne-tuned
through the controls of the DLP projector. However, a perfect
match is impractical as alignment at the sub-pixel level is hard to
achieve and almost impossible to maintain. To avoid moir
and alignment artifacts associated with even a minor misalignment,
we have deliberately blurred the projector image and compensate
for that blur in the LCD image as described in the following sec-
Fresnel Lens and Diffuser
Card in PC
Figure 4: Top: schematic of the HDR display including the pro-
jector, LCD and optics. Bottom: actual photograph of the display.
Both projector and LCD are driven by a single dual-VGA graphics
Using this conﬁguration, the light output of each pixel of the
HDR display is effectively the result of two modulations, ﬁrst by
the DLP and then by the LCD pixel, along the same optical path.
The upper boundary of the dynamic range results from full trans-
mission of both pixels (i.e. the 255
level on both modulators),
and the lowest boundary from the lowest possible transmission of
both modulators (i.e. the 0
level on both modulators). Since the
DLP has a dynamic range of 800 : 1 and the LCD a dynamic range
of 300 : 1, the theoretical dynamic range of the HDR display is
240, 000 : 1. Imperfections in the optical path introduce noise that
reduces the dynamic range to a measured 54, 000 : 1. The lumi-
nance values matching these boundaries are a result of the bright-
ness of the projector and the transmission of the LCD. In this case,
the Optoma EzPro737 is rated at 1200 Lumens, or approximately
2400 Lumens once the RGB color ﬁlters are removed (since each
ﬁlter for red, green and blue eliminates approximately 2/3 of the
incoming light). The Sharp LCD panel has a measured transmis-
sion of approximately 7.6% in the white state (this is quite high for
an LCD since even the theoretical maximum for a color LCD with-
out any losses is only 16% due to the light reduction of 50% at the
polarizer and another 66% due to the RGB color ﬁlter). Assuming
that the light emitted by the HDR display is diffused across a solid
angle ω, the maximum luminance is then given by:
where A is the area of the LCD and Φ
is the maximum outgo-
ing ﬂux. In the HDR display prototype, the ﬂux is approximately
182 Lumens (2400 Lumens ×7.6%). The area A is the area of the
15” LCD (697cm
) and the solid angle of diffusion ω is approx-
imately 0.66sr (40
diffusion horizontally, 15
maximum luminance for this particular conﬁguration is then ap-
proximately 3, 956cd/m
. We actually measured a luminance of
Lumens. The theoretical minimum luminance is less
, while our measurements resulted in a value of
. Clearly, a shift of this range toward even higher lumi-
nance values would be possible with a brighter projector or a more
transmissive LCD. Unlike a standard low dynamic range display,
even an order of magnitude increase of the maximum luminance
would not signiﬁcantly reduce the quality of the ‘black’ state since
is still a very satisfying ‘black’, especially if other parts of
the image contain very high luminance values.
Within that luminance range, a very large number of differ-
ent combinations of output settings for the DLP and LCD can be
achieved. If both systems were linear 8-bit devices then the total
number of combinations would be 256
, over 17,000 of which are
distinct. Due to the non-linear gamma of each system, the actual
range of distinct addressable steps is different, but still signiﬁcantly
larger than what is needed to display the 962 JND steps necessary
to provide all visible and distinguishable luminance steps in the
measured luminance range of the system (including all losses) of
to 2, 700cd/m
4.2 Driving the Projector Display
To correctly render HDR images on this display, we need to analyze
the image formation process of the system. Let us assume for the
moment that both the projector and the LCD panel are perfectly
linear, and that both have the same dynamic range. For now, let
us also neglect the blurring of the projector image. Under these
assumptions, we can achieve the target intensity by normalizing the
intensity range of the HDR image to 0 . . . 1, and using the square
root of this normalized intensity to drive both the projector and the
LCD panel. This even split between pixel values on the projector
and the LCD panel is preferable to a scenario where one value is
very large and the other is very small, since quantization artifacts
are relatively larger for small values.
In reality, neither the projector nor the LCD have a linear re-
sponse, and we also need to compensate for the blurring of the pro-
jector image. We do this in the following way: we choose a simple
estimate of what the projector intensity should be, and then simu-
late the effect of response function and blurring. Finally, the pixel
values of the LCD panel are chosen such that they compensate for
Figure 5: Rendering algorithm for the projector-based display.
The complete rendering algorithm then works as follows (also
see Figure 5): we take the square root of the original HDR image
with intensity I (1). The resulting image (2) represents the target
I for the projector. We map these intensities into projec-
tor pixel values by applying the inverse of the projector’s response
(3). The projector now produces an image of intensity
I, except that the image is actually blurred ac-
cording to p
, the projector’s point spread function (PSF). To simu-
late this blurring, we convolve the projector intensities with the PSF
(4) and divide the result out from the original HDR image to get the
target LCD transparency (5). For the ﬁnal pixel values of the LCD,
we apply the inverse of the panel’s response function r
An exposure sequence of the PSF p
is depicted in Figure 6.
Note the vertical lines visible in images with larger exposure times.
These are the RGB subpixels of the LCD panel. In order to speed
up the computation, we do not use the measured PSF directly, but
ﬁt a tensor-product Gaussian to p
. We usually set the focus of the
projector such that the ﬁtted Gaussian has a standard deviation of
about 2-3.5 pixels, so that we can use a 2D separable ﬁlter of width
13 for the convolution.
Figure 6: Point spread function of the projector.
These image processing steps are possible both completely in
software and using recent programmable graphics hardware (graph-
ics processing unit, GPU). The convolution is the most difﬁcult part
of the GPU implementation, but with the separable approximation
it can be readily implemented as a pixel shader on both the latest
ATI and NVIDIA chips. With this approach we achieve frame rates
of 30 frames per second or more on current hardware.
Response (% of maximum)
Response function - Projector display
LCD panel transparency
Figure 7: Response function of both the LCD panel and the DLP
projector in the projector-based display.
Figure 8 shows the results of this image factorization on a por-
tion of Debevec’s Stanford Memorial Church HDR photograph. On
the left side, you can see a grayscale image that corresponds to the
square root of the original intensity values. Convolving that im-
age with the PSF of the projector yields the center image. This
is the predicted image that will be produced by the projector. Fi-
nally, the right image is the color LCD panel image that corrects
for the blurriness of the projector. It is interesting to note that the
LCD panel image is essentially an edge-enhanced image with low
frequency components attenuated or removed. This is particularly
noticeable for the widths of the window frames. Interestingly, the
algorithm that we apply here is very similar in principle to a local
tone mapping operator. This means that our LCD panel image is
almost a tone mapped version of the original HDR image, although
our method is clearly not designed for that purpose.
4.3 Discussion of the Projector Display
The projector based HDR display hardware provides a tool to
present high quality HDR images but has several drawbacks. In ad-
Figure 8: Factoring an HDR photograph for our projector-based
display. Left: square root of the intensity. Center: blurred image
which is predicted to be the image generated by the de-focused pro-
jector. Right: edge enhanced LCD panel image that corrects for the
blurriness of the projector image.
dition to the obvious form factor problem due to the optical length
required by the projector, the power consumption, cost, thermal
management and video bandwidth requirements are high compared
to a conventional display.
High power consumption and the resulting thermal management
requirements are a consequence of the image creation mechanism
inside the projector. Unlike a cathode ray tube (CRT) display, where
light is created only in the regions of the image that are supposed to
be bright, an LCD or DLP projector creates a uniform light distri-
bution that is then modulated by the LCD or DLP mirror chip. The
power consumption of an LCD or DLP projector is thus indepen-
dent of the image and always very high as there has to be enough
light produced by the lamp such that a full screen ‘white’ can be
shown. Combined with the low modulation efﬁciency of the LCD
or DLP this causes the high power consumption. In the HDR dis-
play the situation is worse than in a conventional, single-modulator
display. The lamp of the projector has to emit enough light to allow
a full screen image at the highest possible brightness of the HDR
display. To achieve 10, 000cd/m
on a 15” screen we would need
an outgoing ﬂux of approximately 500 Lumens (see Section 4.1).
Even with a very high transmission LCD this requires at least 5000
Lumens to be emitted from the projector. In the prototype pre-
sented in Section 4.1 the color wheel/ﬁlter of the projector has al-
ready been removed to reduce the losses in the projector but even so
the modulation efﬁciency of the projector is slightly less than 50%.
The lamp thus has to produce in the order of 10, 000 Lumens. Yet,
in almost all HDR images the area that is actually at such a high
brightness of 10, 000cd/m
is very small. In fact, a random selec-
tion of 100 HDR images indicated that average HDR images have
less than 10% of the image content in the high luminance range
) and that the average luminance over all im-
ages was less than 800cd/m
for indoor scenes and 2, 100cd/m
for outdoor scenes. The projector HDR display consequently cre-
ates a factor of between 12.5 and 4.75 too much light at any given
The projector is also the cause of the high bandwidth require-
ments. Even though the image projected by the projector onto the
back of the LCD is blurred, the projector itself is still a high res-
olution display which requires high resolution input data. As a re-
sult, the projector-based HDR display needs a high resolution video
stream going to the LCD and a similar size video stream to the pro-
jector. This creates a requirement for a dual output graphic card and
imposes limits to the frame rate of the display due to the computa-
Finally, the cost of a high brightness projector is very high which
makes this version of the HDR display unsatisfactory for commer-
cial purposes. Cost also presents a barrier to larger screen sizes
as the brightness requirements increase linearly with area, and the
cost curve for projectors is very steep with brightness (while the
step from 2, 000 Lumens needed for a 15” display to a 4, 000 Lu-
mens projector for a 20” display is only twice as high, the next step
from 20” to 40” TV size would require a 15, 000 Lumens projector
at a price that is over 20 times higher than that of a 4, 000 Lumens
Yet, for research applications, the display is a valuable tool. The
high cost is in part due to the use of fully ﬁnished consumer prod-
ucts instead of individual components, but this also makes it possi-
ble to assemble the system without signiﬁcant development of cus-
tom electronics. Since the drawbacks mentioned above in no way
diminish the actual image quality, the projector-based HDR display
provides researchers with a relatively simple to build solution with
very high image quality.
5 System 2: LED-based Display
As seen in the discussion of the projector-based HDR display in
Section 4.3 there are signiﬁcant obstacles to overcome.
To realize the dream of television or computer displays present-
ing images that look indistinguishable from the real world, it is not
sufﬁcient to merely show images with the appropriate luminance
range and resolution; it is also necessary to make a commercially
viable system that achieves these higher quality images within to-
day’s hardware and software infrastructure and market price points.
The version of the HDR display described in this section retains
the high image quality of the projector display and overcomes the
commercialization barriers: power, thermal, cost, form factor and
infrastructure demands on the graphics card.
5.1 Hardware Setup
We have already seen in Section 4.1 that software correction can
compensate for a low resolution of the rear image of the HDR dis-
play. It is important to realize that this correction works perfectly, as
long as the local image contrast does not exceed the dynamic range
of the front modulator. From the psychology theory presented in
Section 3 we can establish the largest size of a rear image pixel. For
this version of the HDR display we have used light emitting diodes
(LED) at the largest possible size allowed by the veiling luminance
effect which has been validated previously through experimental
tests [Seetzen et al. 2003].
A prototype has been build using 12mm Lumiled Luxeon 1 Watt
white LEDs (LXHL- PWO1) on a hexagonal close-packing matrix
where each LED is individually controlled over its entire dynamic
range with 1024 addressable steps. 760 LEDs have been mounted
behind an 18.1” L.G. Philips LCD with a 500 : 1 dynamic range
and 1280 × 1024 resolution. On full screen white/black the max-
imum luminance is measured as 8, 500cd/m
and the minimum
luminance is zero, since all LEDs are off. The minimum lumi-
nance is less than 0.03cd/m
on a checkerboard pattern larger than
20mm. As mentioned in Section 3.2, the Barten model [Barten
1992; Barten 1993] predicts 1139 JNDs for this luminance range.
The system is capable of displaying images at video rates.
5.2 Driving the LED Display
The principal rendering algorithm for the LED-based system is
quite similar to that for the projector-based display, as shown in
Figure 9. The primary difference between the two display systems
from a rendering perspective is that the PSF of an LED has a much
wider support than the one for a pixel of the projector. Also of
importance is the fact that the LEDs are arranged on a hexagonal
grid rather than a rectangular grid. These differences have two con-
sequences. Firstly, because of the wider support of the PSF, it is
advisable to come up with a better way to choose the LED val-
ues. Since the supports of the PSFs for neighboring LEDs overlap,
determining the optimal LED value is essentially a de-convolution
problem, as explained below. Secondly, because of both the hexag-
onal geometry and wider support of the PSF, the convolution (4)
has to be implemented differently.
Figure 9: Rendering algorithm for the LED-based display.
We address the ﬁrst issue by adding an additional stage (2a) to
derive the target intensities I
for every individual LED. To this end
we ﬁrst down-sample the image to the resolution of the LED array,
and then solve for the values, taking the overlap of the PSFs into
account. This is essentially a de-convolution problem, the full solu-
tion of which would require solving a sparse linear equation system
with as many unknowns as there are LEDs. This is not an option for
interactive applications, and furthermore de-convolution algorithms
are known to be numerically unstable. We can approximate a so-
lution with a single Gauss-Seidel iteration over neighboring LED
pixels. This amounts to a local weighted average of neighboring
LED target values, where some of the weights are negative. In our
design, these LED values (2a) can be forwarded directly to the con-
troller electronics, which correct for nonlinearities in the LED re-
sponse in hardware.
As before, we rely on the LCD panel to compensate for any dif-
ferences between the LED values and the target image. To this end,
we need to forward-simulate the low-frequency image (4) gener-
ated by the LED panel in order to derive the LCD pixel values.
We use two different approaches for software and hardware imple-
mentations. With GPUs, we use a splatting approach, and simply
draw screen aligned quadrilaterals with textures of the PSF into the
framebuffer. Alpha blending is used to accumulate the results. In
software, we approximate the measured PSF with a Gaussian, and
implement the reconstruction by convolving a low resolution image
with the approximate PSF, and then up-sampling to the full resolu-
Response (% of maximum)
Response function - LED display
LCD panel transparency
Figure 10: Response function of the LCD panel in the LED-based
display. The response of the individual LEDs (not shown here) is
designed to be linear.
Figure 11 shows the factorization of an HDR image into LED
component and LCD panel component. Due to the wider support
of the PSF of the LEDs compared to the PSF of a projector pixel in
the ﬁrst setup, the LED image is even more low-pass ﬁltered than
before. As a result, the compensation performed in the LCD panel
is more signiﬁcant, and is visible in the right image.
Figure 11: Factorization of an HDR photograph for the LED-based
display. Left: LED contribution as a result of convolving LED val-
ues with the LED point spread function. Right: edge enhanced
LCD panel image that corrects for the blurriness of the LED image.
5.3 Discussion of the LED Display
The use of LEDs overcomes the power consumption and thermal
problem as a result of intelligent light production, and it eliminates
other commercialization barriers outlined in Section 4.3. Form fac-
tor is no longer an issue as the LED matrix can be the same thick-
ness as a conventional LCD backlight. The video bandwidth re-
quirements are dramatically reduced due to the lower resolution of
the LED matrix. As a result, only a few hundred 8-bit values are
needed over a standard image, and these are added to the front of
the DVI video stream going to the LCD and stripped off by the
controller hardware inside the display. Cost remains an issue but
signiﬁcantly less so than with the projector system.
One slight disadvantage is that the rendering algorithm for the
LED-based display is computationally more demanding than the al-
gorithm for the projector-based display due to the larger support of
the PSF. On a GeForce FX we currently achieve about 10 fps. for a
fullscreen image factorization into LCD and LED components. The
rendering time is mostly limited by having to work around limita-
tions in the support of ﬂoating point framebuffers and textures on
current GPUs. The next generation of GPUs will support a more
complete set of operations, including blending and bi-linear inter-
polation, which should improve the performance of our algorithm
by a factor of 3-4.
We have developed four simple applications to test our display tech-
nology, and to demonstrate its potential in a number of application
domains. The ﬁrst one is a simple HDR image viewer that works
with both displays (Figure 12, top left). It allows the user to load an
HDR image and show it while interactively adjusting the exposure
settings (i.e. the absolute scale of intensity). This application was
also used to generate the images in Figure 1. The three images to
the right of that ﬁgure show a color-coded comparison of an origi-
nal radiance map, an HDR photograph taken off our projector-based
display, and ﬁnally a photograph taken off a standard LCD screen.
The intensities of the photograph from the HDR display are similar
but not identical to the values in the original radiance map. The
differences are mostly due to imperfections in both the calibration
of the display to absolute intensities, and in the image acquisition
process. Clearly both intensity and dynamic range of our display
are vastly superior to the standard monitor.
Figure 12: Screen photographs of the different applications we implemented. The exposure times of the two images in each pair differ by
4 stops. Top left: HDR image viewer. Top right: interactive rendering of measured BRDFs. Bottom left: volume rendering. Bottom right:
medical image viewer.
The second application we developed is related to interactive
photorealistic rendering. We modiﬁed a DirectX application for
displaying BRDFs measured with linear light source reﬂectome-
try [Gardner et al. 2003], and replaced its tone mapping step with
the rendering algorithm for our display (Figure 12, top right). Other
interactive applications that can render into ﬂoating point buffers
can easily be modiﬁed in a similar fashion. As pointed out in
Section 5.2, the rendering performance on current GPUs is limited
due to missing features such as alpha blending with ﬂoating point
framebuffers and lack of bi-linear interpolation for ﬂoating point
textures. These issues should be resolved in upcoming generations
Based on a similar principle, we have built a simple volume ray-
caster that runs on a GPU (Figure 12, bottom left). It allows for
operations such as rotation, slicing, and adjustments to the transfer
function. Both the actual volume rendering algorithm and the pro-
cessing for our displays is implemented in a single Cg shader [Mark
et al. 2003].
Finally, a simple medical imaging viewer and browser has been
implemented to allow radiologist to view medical images in high
dynamic range (Figure 12, bottom right).
7 Conclusions and Future Work
In this paper we have presented two designs for HDR displays,
one based on a projector setup and one based on an LED array.
The two displays we developed have dynamic ranges of well be-
yond 50, 000 : 1, and a maximum intensity of 2700cd/m
, respectively. We have described both hardware and
software aspects of these display systems, and have described a
number of applications for this kind of technology.
There are plenty of opportunities for future work. On the hard-
ware side, the image quality of the LED-based HDR display can
be further enhanced by replacing each white LED with a triplet
of red, green and blue LEDs or a larger number of colored LEDs
if more primaries are desired (e.g. red, blue, blue-green, yellow-
green). Color LEDs have very sharply deﬁned spectral emission
pattern as opposed to the broad band emission of ﬂuorescent tubes
used in conventional LCD backlights. This creates pure primaries
and allows for a signiﬁcant increase in the display color gamut. A
conventional LCD achieves approximately 66% of the NTSC color
gamut while a tri-color LED based system can achieve 98% with
additional red and blue outside the NTSC gamut [Ohtsuki et al.
The actively controlled matrix of LEDs can also be used to very
easily implement backlighting schemes proposed by the industry
such as ﬂashing the backlight in segment in sync with the LCD re-
fresh to reduce motion blur artifacts [Fisekovic et al. 2001]. Like-
wise, the LED array simpliﬁes problems of global luminance non-
uniformity found in conventional displays resulting from a non-
uniform light output of the ﬂuorescent tube backlight, lifetime prob-
lems due to backlight failure, and many other limitations of a con-
ventional LCD backlight.
At this point the calculation of the rear and front image layer are
executed by the graphics card but standardized calculations of this
kind can easily be done by dedicated hardware inside the display
as soon as standards for ﬂoating point video signals (such as exten-
sions to the DVI standard) become available. This will reduce the
load on the graphic card, allow software to directly send ﬂoating
point data to the display, and have the display handle the necessary
computations for blur correction.
In addition to these directions for further developing the hard-
ware, there are a number of potential applications of this display
technology in computer graphics, visualization, human-computer
interaction, and perception research. Commercial applications in-
clude prooﬁng for the ﬁlm and special effects industries, since the
dynamic range of ﬁlm signiﬁcantly exceeds what can currently be
shown on standard displays, and eventually even home theater. For
volume rendering in particular, the higher dynamic range might
open up possibilities for completely new rendering algorithms. We
also plan to explore new interaction metaphors, such as use of HDR
imaging in user interfaces. The extra brightness should be useful
for directing the user’s attention to important events. Finally, an
interesting research direction is to expand the work on human per-
ception research in order to challenge and expand on assumptions
that have been made in the context of traditional display technology.
Other researchers have already used our display to run perceptual
comparisons of tone mapping operators.
We would like to thank Chris Tchou for his work on the initial in-
teractive rendering demo for the linear light source reﬂectometry
data [Gardner et al. 2003]. Paul Debevec made the Daguerreotype
reﬂectance data from the same project available to us (see Figure 12
and the video). The HDR photographs used throughout this paper
were also provided by Paul Debevec, while the CT data set was
provided by Klaus Engel. Thanks to the anonymous reviewers for
their valuable comments.
This work was in part supported by NSERC, IRIS/Precarn, and
the BC Advanced Systems Institute.
ASH IKH MIN , M. 2002. A tone mapping algorithm for high contrast
images. In Proc. of Eurographics Workshop on Rendering 2002,
BA RTEN , P. 1992. Physical model for the contrast sensitivity of
the human eye. In Proc. SPIE, vol. 1666, 57–72.
BA RTEN , P. 1993. Spatio-temporal model for the contrast sensi-
tivity of the human eye and its temporal aspects. In Proc. SPIE,
CHI U, K., HER F, M., SHIRLEY, P., SWAMY, S., WAN G, C., AND
ZIM MER MAN , K. 1993. Spatially nonuniform scaling functions
for high contrast images. In Proc. Graphics Interface ’93, 245–
CHO UDH URY, P., A ND TUMBLIN, J. 2003. The trilateral ﬁlter for
high contrast images and meshes. In Proc. of the Eurographics
Symposium on Rendering, 186–196.
DEB EVE C, P., AN D MA LIK , J. 1997. Recovering high dynamic
range radiance maps from photographs. In Proc. of ACM SIG-
GRAPH ’97, 369–378.
DICOM. 2001. Digital Imaging and Communications in Medicine
ch. Part 14: Grayscale Standard Display Function.
DUR AND , F., AND DORSEY, J. 2002. Fast bilateral ﬁltering for
the display of high-dynamic-range images. ACM Trans. Graph.
(special issue SIGGRAPH 2002) 21, 3, 257–266.
FATTAL, R., LISC HIN SKI , D., AND M.WERMAN. 2002. Gradient
domain high dynamic range compression. ACM Trans. Graph.
(special issue SIGGRAPH 2002) 21, 3, 249–256.
FIS EKOVI C, N., NAU TA, T., CORN ELISSEN, H., A ND BRUI NIN K,
J. 2001. Improved motion-picture quality of AM-LCDs using
scanning backlight. In IDW 2001 Proceedings, 1637–1640.
GAR DNE R, A., TCHOU, C., HAWKI NS, T., A ND DEB EVE C, P.
2003. Linear light source reﬂectometry. ACM Trans. Graph.
(special issue SIGGRAPH 2003) 22, 3, 749–758.
KAN G, S. B., UYT TEN DAE LE, M., WINDER, S., A ND SZELI SKI ,
R. 2003. High dynamic range video. ACM Trans. Graph. (spe-
cial issue SIGGRAPH 2003) 22, 3, 319–325.
LAR SON , G. W., RUSH MEI ER, H., AN D PIATKO , C. 1997.
A visibility matching tone reproduction operator for high dy-
namic range scenes. IEEE Trans. on Visualization and Computer
Graphics 3, 4, 291–306.
MAN N, S., AN D PICARD , R. 1994. Being ’undigital’ with digi-
tal cameras: Extending dynamic range by combining differently
exposed pictures. Tech. Rep. 323, M.I.T. Media Lab Perceptual
Computing Section. Also appears, IS&T’s 48th annual confer-
ence, Cambridge, MA, May 1995.
MAR K, W. R., GL ANVILLE, R. S., AK ELE Y, K., AND KILGARD,
M. J. 2003. Cg: A system for programming graphics hard-
ware in a C-like language. ACM Trans. Graph. (special issue
SIGGRAPH 2003) 22, 3, 896–907.
MIT SUNAGA, T., AN D NAYAR , S. K. 1999. Radiometric self
calibration. In Proc. of IEEE CVPR, 472–479.
MOO N, P., AND SPENCER, D. 1945. The visual effect of non-
uniform surrounds. Journal of the Optical Society of America
35, 3, 233–248.
MUK A, E., AND REIK ER, G. 2002. Reconsidering bit depth for
radiological images – is eight enough? In Proc. SPIE, vol. 4686,
OHT SUK I, H., NAK ANI SHI , K., MO RI, A., SA KAI , S., YACHI ,
S., AND TIMMERS, W. 2002. 18.1-inch XGA TFT-LCD with
wide color reproduction using high power led-backlighting. In
Proc. Society for Information Display International Symposium,
PATTA NA IK, S. N., FERWER DA, J. A., FAIR CHILD, M. D., AND
GRE ENB ERG , D. 1998. A multiscale model of adaptation and
spatial vision for realistic image display. In Proc. of ACM SIG-
GRAPH ’98, 287–298.
REI NHA RD, E., STARK , M., SHI RLE Y, P., AN D FERWERDA, J.
2002. Photographic tone reproduction for digital images. ACM
Trans. Graph. (special issue SIGGRAPH 2002) 21, 3, 267–276.
ROB ERTS ON, M., BOR MAN, S., AND STE VEN SON, R. 1999.
Dynamic range improvements through multiple exposures. In
Proc. of International Conference on Image Processing (ICIP)
SCH LIC K, C. 1994. Quantization techniques for visualization of
high dynamic range pictures. In Proc. of Eurographics Workshop
on Rendering ’94, 7–20.
SEE TZE N, H., WHITEHEAD , L., AND WA RD, G. 2003. A high dy-
namic range display using low and high resolution modulators.
In Society for Information Display Internatiational Symposium
Digest of Technical Papers, 1450–1453.
TUM BLI N, J., AN D RUSH MEI ER, H. 1993. Tone reproduction for
realistic images. IEEE Computer Graphics and Applications 13,
TUM BLI N, J., A ND TURK, G. 1999. LCIS: A boundary hierar-
chy for detail-preserving contrast reduction. In Proc. of ACM
SIGGRAPH ’99, 83–90.
VOS , J. 1984. Disability glare - a state of the art report. CIE
Journal 3, 2, 39–53.