A Wide Field, High Dynamic Range,
Stereographic Viewer
Greg Ward
Exponent – Failure Analysis Assoc.
Menlo Park, California
Abstract
We present a high dynamic range viewer based on the 120-
degree field-of-view LEEP stereo optics used in the original
NASA virtual reality systems. By combining these optics
with an intense backlighting system (20 Kcd/m
2
) and layered
transparencies, we are able to reproduce the absolute
luminance levels and full dynamic range of almost any
visual environment. This technology may enable
researchers to conduct controlled experiments in visual
contrast, chromatic adaptation, and disability and discomfort
glare without the usual limitations of dynamic range and
field of view imposed by conventional CRT display
systems. In this paper, we describe the basic system and
techniques used to produce the transparency layers from a
high dynamic range rendering or scene capture. We further
present an empirical validation demonstrating device's ability
to reproduce visual percepts, and compare this to results
obtained using direct viewing and a visibility matching tone
reproduction operator presented on a conventional CRT
display.
Introduction
The natural world presents our visual system with a wide
range of colors and intensities. A starlit night has an
average luminance level of around 10
-3
candelas/m
2
, and
daylight scenes are close to 10
5
cd/m
2
. Humans can see
detail in regions that vary by 1:10
4
at any given adaptation
level, over which the eye gets swamped by stray light (i.e.,
disability glare) and details are lost. Modern camera lenses,
even with their clean-room construction and coated optics,
cannot rival human vision when it comes to low flare and
absence of multiple paths (“sun dogs”) in harsh lighting
environments. Even if they could, conventional negative
film cannot capture much more range than this, and most
digital image formats do not even come close. With the
possible exception of cinema, there has been little push for
achieving greater dynamic range in the image capture stage,
because common displays and viewing environments limit
the range of what can be presented to about two orders of
magnitude between minimum and maximum luminance. A
well-designed CRT monitor may do slightly better than this
in a darkened room, but the maximum display luminance is
only around 100 cd/m
2
, which doesn’t begin to approach
daylight levels. A high-quality xenon film projector may
get a few times brighter than this, but they are still two
orders of magnitude away from the optimal light level for
human acuity and color perception.
In this paper, we present a novel device for displaying
stereographic, high dynamic-range images to a single
viewer. This static display device combines simple, known
technologies in a way that faithfully reproduces the visual
field of a natural or simulated scene, resulting in a
heightened sense of realism and presence. Following a
description of the device and the requisite image processing
techniques, we demonstrate the quality of its output by
comparing visibility in the original scene to visibility in
our HDR reproduction. We conclude with some ideas for
future development of this technology in cooperation with
Bristol University in England, and the University of British
Columbia in Vancouver.
Device Description
The high dynamic-range viewer itself is a simple device,
consisting of a bright, uniform backlight joined to a set of
LEEP ARV-1 optics. The optics allow a 120° field of view
in each eye for a very natural and complete stereo view from
a pair of 2.5” (6.4 cm) fisheye images mounted side-by-side.
A schematic of the viewer is shown in Fig.!1, and the
actual viewer is shown in Fig.!2. Four viewers have been
constructed thus far, and this is the latest model.
Figure 1. Schematic of HDR viewer.
12 volt 50 watt lamps
heat-absorbing glass
reflectors for uniformity
diffuser
ARV-1 optics
cooling fan
Figure2. HDR viewer device.
The transparencies sit on top of the diffuser in front of the
ARV-1 optics. Focus adjustment is performed by shifting
the optics slightly closer or away from the transparencies,
which are held in place by a clip. Precise focus is not
important, as the eye will accommodate over a certain range.
It is important when viewing to remove eyeglasses that
might prevent the eyes from getting close enough to the
optics to provide the full field of view. Provided the wearer
does not have severe astigmatism, the focus adjustment
permits acuity at the visual limit of 20 cycles/degree,
provided the transparency is rendered at a density of 800 dpi
or better (2048¥2048 resolution).
Image Processing
The basic transformation required for correct perspective
viewing with the ARV-1 optics is a hemispherical fisheye
projection. This projection is most easily explained in the
form of a diagram, as shown in Fig.!3.
Figure 3. The hemispherical fisheye projection.
In a hemispherical fisheye projection, the distance from the
center of the image is equal to the sine of the angle from the
principal view axis (i.e., depth). This can be visualized as a
hemisphere placed on the image plane and projected directly
down onto it, as shown if Fig.!3. Each pixel in the image
corresponds to the ray that passes through the point in the
hemisphere directly above it. In vector terms, this ray is
defined by the origin view point and the unit direction vector
given in Eq.!1. The x and y values in the equation are the
offsets from the image center normalized to half the width of
a full 180° image. These coordinates will equal 1 or –1 at
the image edges, and 0 in the center. The unit view vectors
for x, y, and z correspond respectively to the horizontal,
vertical, and depth directions for this image. Note that the
view is not defined if (x
2
+ y
2
) > 1.
†
ˆ
v
p
=
ˆ
v
x
x +
ˆ
v
y
y +
ˆ
v
z
1- x
2
- y
2
(1)
Since each image covers only 120° rather than 180°, the
only the corners of the image are perpendicular to the
principal view axis. The final projection can be seen in the
processed image shown in Fig.!4.
Figure 4. A detail transparency layer for the HDR viewer.
Due to chromatic aberration in the LEEP ARV-1 optics, it
is best to precorrect the image by scaling the red image
plane proportionally more than the blue image plane so that
the red is about 1.5% larger than the blue, and the green is
in between. This can be seen as a red fringe near the edges
of Fig.!4, which results in fairly good color convergence
when viewed through the LEEP optics. The middle of the
image not subject to the same chromatic distortion, so the
center of view is unaffected.
Due to limitations in the film recording process, it is
not possible to achieve the full, desired dynamic range in a
single transparency. Although the film recorder we used is
capable of producing nearly 1000:1 at the limit, the bottom
reaches of this range have fairly large intensity steps. The
useful range where the intensity steps are below the visible
view point
pixel
pixel direction
q
threshold necessary to avoid banding artifacts is closer to
100:1. Since we wish to produce images with a dynamic
range in excess of 10,000:1, we need to use two
transparencies, layered one on top of the other. By
combining the ranges in this way, we double our dynamic
range in log terms as the densities add together.
To avoid problems with transparency alignment and
ghosting, we reduce the resolution of one layer using a
Gaussian blur function. Because the dynamic range of the
individual color channels is not important for perception, we
convert to gray in the scaling layer to simplify our
calculations. The resulting image is shown in Fig.!5. The
degree of blur is not critical – we use a blur radius equivalent
to 16 cycles across the image. We have found this to allow
for easy image alignment without seriously compromising
the maximum luminance gradient in the combined result.
Figure 5. A scaling transparency layer for the HDR viewer.
Since the combination of the two transparency layers is
equivalent to adding the densities, each image must have half
the density of the original. This is easily accomplished by
taking the square root of each pixel value in the blurred
image, and dividing the original image by this image for the
detail layer. This result is the one shown in Fig.!4. The
effects of dividing by the scaling layer may be seen in the
odd gradients near the window’s edge, which disappear when
the layers are placed together in the HDR viewer. Alignment
marks at the corners of the images aid in their registration.
So far, we have discussed only the generation of a
single layered image, whereas two images are required for a
stereo view, one for each eye. In cases where stereo
perspective is negligible or unimportant, a single image
may be duplicated for each eye to achieve a satisfactory
monocular view. For stereoscopic viewing, one must
capture or render two images, and this is done by simply
shifting the real or virtual camera by the average interocular
distance, which is approximately 2.5” (6.4 cm). It is not
necessary to adjust the principal view axis in the two
perspectives, as the observer should and will make his or her
own accommodation when viewing the stereo pair.
Validation Experiments
We have performed two types of validation for the HDR
viewer’s performance, one quantitative and one qualitative.
In the quantitative study, we wanted to measure the
luminances produced by the viewer and compared them to
the original input to determine that the method of
combining transparency layers performs as specified. In the
qualitative study, we took several subjects into a darkened
room and presented them with a contrast visibility chart,
which we then reproduced in our HDR viewer. We start by
discussing the quantitative study, and follow with a
presentation of the qualitative results.
Quantitative Study
The quantitative measurement process presented some
challenges, as there are no luminance probes with sufficient
spatial resolution and freedom from stray light to measure
the very large gradients we produce in the viewer. We
therefore restricted our quantitative validation to a simple
verification that combining transparencies adds densities as
predicted. To accomplish this, we used an industry standard
densitometer made by X-rite, and found that layering
transparencies corresponds to adding densities within the 2%
accuracy of the densitometer. Mathematically, the greatest
theoretical error will happen near the lowest densities, where
multiple interreflections could increase the effective
transmittance by as much as 1%. This is not enough of an
error to be significant in most simulations.
Figure 6. A contrast chart designed by Tom Ayres [Ayres96].
Qualitative Study
In our qualitative comparison, we asked subjects to view the
chart shown in Fig.!6 in a darkened room, illuminated
dimly by a single, distant spotlight designed to simulate a
car’s headlights. We then photographed the same chart
using an Olympus E–10 digital camera and a method for
combining multiple exposures into a calibrated, high
dynamic-range image [Matsunaga&Nayer99]. This
calibrated image is shown in Fig.!7.
Figure 7. A high dynamic-range photograph of the contrast
chart under the initial viewing conditions.
Unfortunately, the resolution of the photograph in Fig.!7 is
not very high, and this turned out to be a problem for our
tests. Although we could have used a longer focal length on
our camera and thus captured better resolution in the chart
image, we needed to capture as much of the surround as
possible in order to have proper adaptation in the HDR
viewer. The results for the smaller disk visibilities should
not be taken too seriously for this reason.
In a second test, we shone a bright spotlight directly in
the subjects’ eyes directly under the chart to simulate a glare
condition akin to a car’s oncoming headlights and asked for
which disks they could discern the orientation. This same
condition was reproduced in the HDR viewer by introducing
a white disk to the transparency with the appropriate size and
position to produce the same effective luminous power on
the retina as in the test condition.
The results of these two test conditions are shown for
two subjects in Table!1 for the original scene and the HDR
viewer. The viewing results give the number of large and
small disks discernable, respectively. (In our tests, the
visibility ordering of the disks never changes, so there is no
need to account for the individual disk visibilities.)
Table 1. Disk Visibility: Real vs. HDR viewer.
Subject
Condition
Real
HDR viewer
GW
normal
5, 2
5, 0
glare
2, 0
2, 0
CS
normal
6, 2
5, 0
glare
3, 0
3, 0
As we can see in our results, the smaller disks were never
quite visible in the HDR viewer, and we believe this is due
to the limited resolution of the original photograph rather
than a limitation of the viewer itself. We would like to
repeat this experiment using a higher resolution image from
a line-scan panoramic camera, but at the time of this
writing, we have not yet managed to do so. On the other
hand, the large disk visibility was reproduced fairly well in
our experiment, indicating that the basic technique of
recreating luminances in a wide-field stereo viewer is an
effective means of virtual reenactment in a luminance range
corresponding to roadway illumination conditions.
One problem we noticed with reproducing the glare
condition was that the interreflections in the HDR viewer
optics were a source of difficulty. In particular, the bright
region corresponding to the oncoming headlight reappeared
as a ghost in the view, obscuring other parts of the image in
the same way that “sun dogs” appear in a photograph
containing the solar disk. One way to alleviate this problem
is to coat the optics to reduce reflections, and this may be a
desirable enhancement to future versions of the viewer. It is
also helpful to use the viewer in a darkened room if the
scene being reproduced is very dim, as stray light can
otherwise enter from the sides and obscure the view and
adversely affect viewer adaptation.
For comparison purposes, Fig.!8 shows the scene as
simulated by the visibility-matching tone reproduction
operator of Larson et al. [Larson97]. Fig.!8a shows the
chart under the “normal” condition, and Fig.!8b shows the
chart under the glare condition. Unless printing is carefully
controlled, there is no guarantee that the print you see in the
proceedings will reproduce the target visibility, but it
worked reasonably well on our monitor.
Figure 8a. A visibility-matching tone operator applied to the
normal chart viewing condition.
Conclusions
We have presented a simple device for displaying high
dynamic-range, wide-field stereo imagery. The most serious
limitation of the current device is its restriction to static
scenes. Clearly, it would be better if we could quickly
change from one image to another, without needing to swap
transparencies and interrupt the observer. Ideally, we would
like to present animations of virtual reenactments or
reproductions at video frame rates. The challenge of
achieving the necessary resolution and bandwidth, although
significant, may be met with today’s PC graphics hardware.
It may require multiple cards in the same machine and
specialized drivers, but it is possible.
A group of researchers in the physics department at the
University of British Columbia in Vancouver under the
direction of Lorne Whitehead have already constructed a
prototype high dynamic-range display with animation
capabilities. Although their design is currently restricted to
a typical monitor’s field of view, there is good reason to
believe that the same technology could also be applied in a
wide-field stereo configuration. Another research group in
the computer science department of Bristol University in
England is interested in doing just that, and the author hopes
to collaborate with both groups to achieve these goals.
References
1. Thomas Ayres, Psychophysical Validation of
Photographic Representations, ASME 1996.
2. Greg Ward Larson, Holly Rushmeier, Christine Piatko, A
Visibility Matching Tone Reproduction Operator for High
Dynamic Range Scenes, IEEE Transactions on
Visualization and Computer Graphics, Vol. 3, No. 4,
December 1997.
3. T. Mitsunaga and S. K. Nayar, Radiometric Self
Calibration, Proc. IEEE Conference on Computer Vision
and Pattern Recognition, June 1999.
Biography
Greg Ward (a.k.a. Greg Ward Larson) graduated in Physics
from UC Berkeley in 1983 and earned a Master’s in
Computer Science from SF State University in 1985. Since
1985, he has worked in the field of light measurement,
simulation, and rendering variously at the Berkeley National
Lab, EPFL Switzerland, Silicon Graphics Inc., Shutterfly,
and Exponent. He is author of the widely used Radiance
package for lighting simulation and rendering.
![Figure 6. A contrast chart designed by Tom Ayres [Ayres96].](/figures/figure-6-a-contrast-chart-designed-by-tom-ayres-ayres96-a5lnms4f.webp)
