TauBench: Dynamic Benchmark for Graphics Rendering
Joel Alanko
, Markku M
and Pekka J
Tampere University, Finland
Rendering, Graphics File Formats, Virtual Reality, Animation.
Many graphics rendering algorithms used in both real time games and virtual reality applications can get
performance boosts by reusing previous computations. However, the temporal reuse based algorithms are typ-
ically measured using trivial benchmarks with very limited dynamic features. To this end, we present two new
benchmarks that stress temporal reuse algorithms: EternalValleyVR and EternalValleyFPS. These datasets
represent scenarios that are common contexts for temporal methods: EternalValleyFPS represents a typical in-
teractive multiplayer game scenario with dynamically changing lighting conditions and geometry animations.
EternalValleyVR adds rapid camera motion caused by head-mounted displays popular with virtual reality ap-
plications. In order to systematically assess the quality of the proposed benchmarks in reuse algorithm stress
testing, we identify common input features used in state-of-the-art reuse algorithms and propose metrics that
quantify changes in the temporally interesting features. Cameras in the proposed benchmarks rotate on av-
erage 18.5× more per frame compared to the popular NVidia ORCA datasets, which results in 51× more
pixels introduced each frame. In addition to the camera activity, we compare the number of low confidence
pixels. We show that the proposed datasets have 1.6× less pixel reuse opportunities by changes in pixels’
world positions, and 3.5× higher direct radiance discard rate.
A pleasant and interactive virtual 3D experience re-
quires the display to update a new image in high fre-
quency, but rendering a realistic image takes time.
However, the next frame is usually very coherent with
the previous one (Yang et al., 2009), even when the
rendered content is very dynamic. This coherency
can be utilized in order to decrease the computational
effort of rendering. These are called temporal reuse
methods, which means that the previously rendered
image is used in some way to accelerate the computa-
tion to render a new one.
Often when the performance of methods and pro-
cesses is compared, benchmarks are created and used.
Benchmarks contain reproducible test scenarios that
are used as an input for algorithms. The results can
then be compared with the confidence that the test was
performed in a fair setting.
For temporal reuse algorithms, a benchmarking
setting would be a dataset that contains 3D data and
animations required in the image rendering. It would
be easier to compare algorithm development advance-
ments with standard benchmarks, having access to
previously understood and used dynamic datasets.
Moreover, such a benchmark would benefit the field
of temporal rendering by showing how and where the
state-of-the-art algorithms succeed and fail in render-
ing high-quality animations. It would also serve as a
challenge to motivate pushing rendering development
However, there are very few such datasets released
in public, and graphics research rarely uses them.
There are at least two obvious reasons for this. First,
authors create the datasets themselves, own an IP they
can use, or buy a set with animations that cannot be
released to the public. When such datasets are only
present in their research papers, it serves as a potential
bias towards the novelties the researchers are propos-
ing, as it is impossible to reproduce the same case.
Furthermore, gathering and creating these datasets
takes time and effort, and polishing them to release
quality would increase it even more (Tamstorf and
Pritchett, 2019). This significant time investment to
develop datasets tends to be avoided, resulting in the
datasets having uninteresting animations and raising
the bar to release them. Second, because very few
Alanko, J., Mäkitalo, M. and Jääskeläinen, P.
TauBench: Dynamic Benchmark for Graphics Rendering.
DOI: 10.5220/0010819200003124
In Proceedings of the 17th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2022) - Volume 1: GRAPP, pages
ISBN: 978-989-758-555-5; ISSN: 2184-4321
2022 by SCITEPRESS Science and Technology Publications, Lda. All rights reserved
datasets have been released, in varying file formats,
there is no single clear dataset format to select from
because there are plenty of standard file formats used
across the industry. In summary, the academic re-
search papers use datasets with animations to produce
convincing results, but the datasets are rarely publicly
available or released to the public.
In this paper, we propose TauBench, which con-
sists of two new dynamic benchmarks that provide
a challenge for the temporal reuse methods. These
LEYFPS are the first publicly released dynamic
benchmarks with a permissive license (CC-BY-NC-
SA 4.0) containing interactive rendering use contexts.
The animations in the benchmark dataset are more
challenging for temporal rendering than any previ-
ously published datasets. We evaluate this through
various metrics, comparing the fundamental temporal
properties that most reuse methods utilize.
Our main contributions in this paper are the fol-
We publish a new dynamic benchmark TauBench
with two datasets: ETERNALVALLEYVR: a VR
camera benchmark representing a realistic inter-
active VR use case, and ETERNALVALLEYFPS:
a fast-paced camera benchmark that represents
the typical interactive first-person application use
case. The datasets are available at https://
We present comparisons on the temporal aspects
of the TauBench benchmarks and show that both
datasets have a significantly higher motion with
the camera, lights, skeletal and rigid body objects
in the animation. Comparisons are made by mea-
suring the positional and rotational camera’s ac-
tivity, the number of new pixels appearing to the
camera’s frustum, and the change in the features
used by temporal reuse methods.
2.1 Temporal Coherency
Techniques that utilize frame-to-frame coherency are
standard practices across the graphics rendering in-
dustry (Yang et al., 2020). However, there have
not been temporal complexity comparisons made be-
tween different 3D dataset animations to the best of
our knowledge, but temporal coherence has been uti-
lized with different input features. We highlight a few
families of techniques where dynamic benchmarks
can be particularly beneficial, namely anti-aliasing,
upsampling, asynchronous reprojection used in vir-
tual reality applications, solving the path tracing inte-
gral with low sample count reconstruction using de-
noising, and deep learning reconstruction. Tempo-
ral coherency methods are discussed more thoroughly
in (Scherzer et al., 2012).
Mueller et al. introduced a temporal change met-
ric that detects when colors differ by a perceptually
noticeable amount (Mueller et al., 2021). They com-
pare RGB values of tone mapped frames with a set
of thresholds, and show with user studies when the
pixels differ too much to be under the just-noticeable-
difference limit. We use this color change threshold
metric in our comparisons.
To detect geometry edges and adapt the denois-
ing filter based on them, McCool proposed compar-
ing adjacent pixels’ normalized color distance, world
space position difference, and normal orientation dif-
ferences (McCool, 1999). Since then, similar denois-
ers have been utilizing additional feature buffers con-
taining base color, shading normal, world space posi-
tion, and direct and indirect radiance (Hanika et al.,
2011; Li et al., 2012; Sen and Darabi, 2012; Kalantari
et al., 2015). The feature comparisons presented in
(McCool, 1999) also serve as a basis for some of our
comparison metrics.
Temporally caching radiance and computation-
ally taxing irradiance has also seen reuse techniques
(Ward and Heckbert, 1992; Kriv
anek et al., 2005;
Tawara et al., 2004), which utilize extra features ra-
diance, irradiance, depth and object identifiers.
Converging the radiance and irradiance to the fi-
nal image after a short period of time is used in up-
dating irradiance light probes (Gilabert and Stefanov,
2012; Majercik et al., 2019) and dynamic virtual point
lights (Keller, 1997; Wald et al., 2003; Hedman et al.,
2017; Tatzgern et al., 2020). We use the direct and
indirect radiance in our dataset comparison metrics.
2.2 Published Dynamic Datasets
NVidia Open Research Content Archive (ORCA) is
a professionally-created 3D assets library from 2017
that has been openly released to the research com-
munity (Amazon Lumberyard, 2017; Nicholas Hull
and Benty, 2017). The Bistro datasets were created
to demonstrate new anti-aliasing and transparency
features of the Amazon Lumberyard engine and the
Emerald Square dataset to go along with the re-
lease of research renderer Falcor (Benty et al., 2020;
Amazon.com, Inc., 2021). There is also a dataset
by “Beeple” called Zero Day that has a high count
of dynamically changing scenery and emissive tri-
angles (Winkelmann, 2019). All the files are in
TauBench: Dynamic Benchmark for Graphics Rendering
FBX file format, and they contain camera anima-
tions, modern textures, and modern geometry com-
plexity. The datasets run for 60–100 seconds, and
their animated cameras have 11–17 key frames de-
scribed. The ORCA datasets have the most modern
geometrical and material representation, with over a
million surface faces and physically-based materials,
and they feature an animated camera that flies through
the dataset. These features make them the current
state of the art in dynamic benchmarks. None of these
datasets were released for benchmarking purposes in
mind, but they are great examples of what kind of
datasets are often used by the research community
when investigating temporal reuse.
A Benchmark for Animated Ray Tracing (BART)
was released in 2001 (Lext et al., 2001). It has three
datasets: Kitchen, Robots, and Museum. All of the
datasets are described with an infrequently used file
format called Animated File Format (AFF), which is
an extension of a file format called Neutral File For-
mat (NFF) (Haines, 1987), providing properties to de-
scribe animations. The test suite has been released
with benchmarking purposes to measure ray tracing
performance and has been used in dynamic ray trac-
ing research. Each dataset is designed with a specific
stress goal in mind. The Kitchen dataset has signifi-
cant differences in the density of the details, memory
cache performance with hierarchical and rigid body
animations, and varying frame-to-frame coherency in
the animations. The Robots dataset focuses on the
hierarchical animation, distribution of objects in the
dataset, and bounding volume overlapping, and the
Museum dataset focuses on the efficiency of ray trac-
ing acceleration structure rebuilding. They also pro-
pose methods to measure and compare errors when
datasets are used with ray tracing algorithms.
The Utah repository collection was released by
Wald in 2001 (Wald, 2019). The datasets were re-
leased along with two research articles focusing on
dynamic ray tracing (Wald et al., 2007; Gribble et al.,
2007). The motivation behind setting up the repos-
itory with the datasets was that ray tracing was be-
coming viable for interactive applications. These sets
are described with a series of OBJ files and use MTLs
to define the material models, and they do not have
camera descriptions released, as they mainly focus on
the ray tracing acceleration aspect with these datasets.
Wavefront OBJ is a human-readable file format to
describe 3D geometry and rendering primitives, and
the MTL file format describes the colors, textures,
and reflection maps (Bourke, 2011; Ramey et al.,
1995). Describing an animation with a series of OBJ
files means that the triangles are animated separately,
which is commonly also called morph targets or key
shapes (Alexa et al., 2000).
Various other released 3D datasets have tempo-
ral aspects. The Moana Island Scene is a complete
animation dataset featured in the 2016 Disney film
Moana (Walt Disney Animation Studios, 2018). The
open source 3D animation short film Sintel has been
used in the creation of the MPI Sintel Flow Dataset
to be used in motion flow algorithm research (Butler
et al., 2012). In addition to Sintel, the Blender Foun-
dation provides plenty of openly available datasets in
their demo files, displaying the new features for their
rasterizer and path tracer (Blender Foundation, 2021).
UNC Dynamic Scene Benchmarks have animations
of breaking objects and non-rigid object deforma-
tions (The GAMMA research group at University of
North Caroline, 2018). Similarly, the KAIST Model
Benchmarks have animated fracturing objects, cloth
simulations, and walking animated characters (Sung-
eui, 2014). The downside of these datasets is the lack
of temporally challenging scenarios. They either have
slowly moving cameras, aged material models, or do
not contain moving lights and objects.
3.1 Camera’s Activity
A virtual reality camera can move with six degrees of
freedom (DOF): the camera can translate in three di-
rections in XYZ coordinate space, and rotate around
three axes with pitch, yaw, and roll. In a typical first-
person application controlled with a mouse, the roll
rotation is restricted, resulting in five DOF. We calcu-
late the distance d
the camera travels each frame:
, p
) =k
k, (1)
where p
is the camera’s position on this frame, and
its position in the previous frame.
We also calculate the amount of rotation d
happens between frames for each of the three axes
) =k θ
k, (2)
where θ
is the angle in pitch, yaw, or roll rotation for
the current frame, and θ
the angle for the previous
Moreover, we determine whether the pixels are
outside of the frustum with a discard function
f rustum
(x,y,w,h) =
1, if (x < 0) k (w 1 < x),
1, if (y < 0) k (h 1 < y),
0, else,
GRAPP 2022 - 17th International Conference on Computer Graphics Theory and Applications
Figure 1: Rendering of the Sponza scene. From left to right: reference orientation, camera turned 25 degrees from the
reference with pitch rotation upwards, yaw rotation to the left, and roll rotation. The checkerboard pattern shows the frustum
discard areas that do not have any valid temporal history data. Quick movements, especially in the yaw direction, rapidly
invalidate all the available history, whereas roll rotation can invalidate only the corners.
where x, y are the reprojected screen space coordi-
nates and w,h are the width and height of the screen.
With f
f rustum
, we form a binary discard mask for the
frame’s pixels. When the camera has rotation from
the previous frame to current in pitch, yaw, or roll ro-
tation, the discarded pixel history is vastly different,
as seen in Figure 1.
We apply the frustum discard function to get the
discarded percentage f
(w,h) of pixels per im-
age by:
(w,h) =
f rustum
where x
are the reprojected coordinates retrieved
with indices i, j running through the size of the im-
age’s width w and height h. Finally, we calculate
the mean of the discarded pixels for the duration of
the animation with
(w,h), where N
is the number of frames in the animation.
3.2 Temporal Feature Comparison
An apparent factor of a temporal challenge is the
frame-to-frame change in the feature buffers used by
temporal reuse algorithms. When comparing these
values for the back reprojected current pixel and the
previous frame’s pixel, the distance between the two
can determine whether the new pixel is coherent with
the previous one. We seek to compare these chal-
lenges between datasets. Hence, we form a general
metric f
, similar to the depth-based edge-detection
estimator by (McCool, 1999), where we compare the
back projected current frame’s feature buffer value B
to the previous frame’s corresponding value B
) =
1, if k
k > a
0, else,
where a
is a confidence threshold value. We use
this metric for computing the distance in world space
positions, shading normals, direct radiance, and in-
direct radiance, with the respective thresholds a
, a
, and a
. Figure 2 illustrates the rele-
vance of the chosen features. The leftmost column
Figure 2: Different masks are composed of the temporal
reuse methods’ input features. On the third row, we display
the difference between the first and second frames. Com-
paring the change in pixels’ world positions, we recognize
(a.) disocclusions and (b.) occlusions. Disoccluded parts
must quickly forget the history buffer, and occluded parts
may be reconstructed using back reprojection. When shad-
ing normals are compared, we recognize too big of a change
in (c.) normal directions between the two frames. The nor-
mal angle has changed so much, and we should validate the
usability of these pixels. On the third column, a light has
moved from right to left, resulting in four different tempo-
rally unstable parts: (d.) new direct light, (e.) previously lit,
(f.) previously shadowed, and (g.) newly shadowed areas.
Finally, on the rightmost column, an appeared wall reflects
diffusely (h.) new indirect light on the scene.
shows how occlusions and disocclusions can be rec-
ognized by comparing the distance in world posi-
tions. Temporal methods try to restart the temporal
history for the disoccluded pixels appearing behind
the sphere, and reproject the pixels occluded by the
sphere with the help of the motion vectors. Mueller et
al. presents in (Mueller et al., 2021) that the temporal
change in pixel colors can stay unnoticed by human
when it changes less than 16/255 with 8 bit RGB col-
ors. Inspired by their work, we tune our thresholds to
32/255. This makes sure we compare pixel changes
that would most likely be noticed, and to mitigate the
effect of each dataset’s geometry being scaled differ-
ently. Running the function f
through the pixels
in a frame yields a mask like the one shown in the
TauBench: Dynamic Benchmark for Graphics Rendering
leftmost column of Figure 2, containing all the pixels
either occluding or disoccluding the geometry. Simi-
larly, columns 2–4 demonstrate the respective masks
obtained with the shading normal metric f
, di-
rect radiance metric f
, and indirect radiance metric
. In particular, the shading history may become
less valid, when the reprojected pixels’ shading nor-
mal angle changes drastically, or when lights change
their position or moving geometry obstructs a shaded
point, or when after a few bounces, the light reaches
places not previously lit. The confidence thresholds
for these metrics are also tuned for each dataset.
4.1 Capturing
Temporally challenging properties are intrinsic as-
pects of games, so we captured the animation datasets
from an open source multiplayer arena shooter
game called Cube 2: Sauerbraten (Oortmerssen,
2021). Sauerbraten was selected for the captur-
ing as it was openly available, contained the ge-
ometry in common triangle format, and the con-
tent was fast paced. We captured two datasets from
different areas of the Eternal Valley map, released
with a permissive Creative Commons Attribution-
NonCommercial-ShareAlike 4.0 Unported License
(CC BY-NC-SA 4.0). The map contains the geome-
try of a sizable outdoor scenery, with the sky directly
casting sunlight to the valley, illuminating half of it.
Moreover, we updated the scenery with modern GGX
materials. The scenery now uses textures with 2K res-
olution in the base color, normal map, metallic map,
and roughness maps. Using the modern open source
3D file format specification glTF 2.0, we encapsulated
the datasets to only two singular files: ETERNALVAL-
LEYFPS and ETERNALVALLEYVR. A representative
selection of rendered frames of the two datasets are
shown in Figure 3.
The first-person camera movement that we cap-
ture in ETERNALVALLEYFPS is one of the most
common modes in interactive games. It aggregates
those essential aspects of interactive scenarios that
produce highly temporally changing rendering set-
tings: rapidly changing camera position and irregu-
larly rotating camera orientation. The quick changes
around the dataset put a burden on the rendering
methods utilizing geometry occlusions and disocclu-
sions, and the camera changes tax the handling of new
pixels revealed outside of its frustum.
We used the Unity game engine with the Oculus
Quest 2 virtual reality headset to capture HMD cam-
era movement for ETERNALVALLEYVR. The move-
ment is unique, as the head turns and rotates quickly
in a way not possible in traditional PC interactive ap-
plications. The head is in constant motion and rotates
around each axis.
We capture the activity of cameras and geome-
try animations 60 times per second, for 6 seconds.
Comparison datasets are overly lengthy: typical fps
selection for real time context is 60, so for example
BISTROEXTERIOR that is 100 seconds long, compar-
ison made against ground truth images would require
rendering over 6000 frames with high sample count.
With 6 seconds we find a balance of understandable
and useful content in the animations for comparisons,
and rendering time of ground truth images.
Both of the datasets, and rendered videos of them,
are available at https://zenodo.org/record/5729574.
4.2 Temporal Measurements and
We render all datasets with Blender’s path tracer Cy-
cles and extract feature buffers, namely world-space
positions and normals, direct and indirect radiance,
separately. Blender’s Cycles is a path tracing renderer
with many supported features, like skeletal anima-
tions. We render all of the features with a 1920×1080
resolution, and the pixels are sampled with 1024 paths
with a maximum of 12 light bounces. Animations are
rendered at 24 fps to guarantee at least some change
even in the most stable datasets. With these config-
urations, we have a reasonable rendering time and a
high enough sample count for the indirect buffer to
converge enough, mitigating most of the path tracing
We compare our benchmark with mod-
ern popular datasets in the ORCA library by
RIOR (Amazon Lumberyard, 2017), and EMERALD-
SQUARE (Nicholas Hull and Benty, 2017). These
sets have not been released as temporal rendering
benchmarks, but they represent well the datasets
often used by the research community, as they have
modern geometrical complexity and physically-based
material models. In addition, we compare with the
TOASTERS dataset from the Utah repository (Wald,
2019), which consists of a vertex morphing setting
that is apparent in most of the previously released
rendering benchmarks (The GAMMA research group
at University of North Caroline, 2018; Sung-eui,
2014). For the comparison we normalized all scenes
to human sized scale.
The animation details of the TauBench datasets
and the comparison datasets are presented in Table 1.
GRAPP 2022 - 17th International Conference on Computer Graphics Theory and Applications
Table 1: Animation details of the datasets.
armatures static
ETERNAL VALLEY FPS 477 32 73 617
ETERNAL VALLEY VR 491 34 63 626
TOASTERS - - - -
Table 2: Change in camera’s position.
average variance max
ETERNAL VALLEY FPS 1.892 0.826 5.475
ETERNAL VALLEY VR 0.041 0.160 7.163
BISTRO INTERIOR 0.013 0.000 0.018
BISTRO EXTERIOR 0.055 0.001 0.143
EMERALD SQUARE 0.154 0.004 0.248
In particular, the proposed datasets have significantly
more dynamic rigid objects and armatures. ETER-
NALVALLEY is a large scene, so most of the changes
may not be affecting the final render, but the pro-
posed datasets still have much more dynamic features
compared to the other sets, as the highest compar-
ison count is on the BISTROEXTERIOR, which has
15 small lamp bulbs swaying slowly by the force of
the wind. Moreover, our datasets have over a thou-
sand dynamic point lights appearing, moving and dis-
appearing throughout the animation. In contrast, the
only dynamic light sources in the comparison datasets
are the animated bulbs in the BISTROEXTERIOR,
which have emissive texture on them. However, in
our case, the scene is under direct sunlight, making
the effect of these bulbs on the final shading negligi-
The amount that the camera translates around
the datasets varies significantly, as seen in Table 2.
The dataset ETERNALVALLEYFPS moves around the
scene the most and has the highest singular per-frame
change in position compared to the other datasets.
Another noticeable aspect in the movement of the pro-
posed datasets is the continuous small changes in its
motion compared to other datasets, which is shown in
the more considerable variance. In contrast, the other
datasets keep the constant value for a while and then
jump abruptly to a new one. This shines some light
on the main difference between the proposed dataset
and the previous work: our dataset’s animation key
frames have been recorded during the game play with
high frequency, whereas the previous work has the
key frames placed by the animator, letting the cam-
era fly between marked points linearly.
The proposed datasets also have a more significant
change in all rotation angles during the camera anima-
tion than any other set, as shown in Table 3. Further-
more, compared to the others, the dataset ETERNAL-
VALLEYVR is the only one with considerable roll ro-
tation. This is explained by it being captured with a
Table 3: Change in camera’s rotation per axis, in degrees
per frame.
1.550 2.839 3.365 8.887 0.000 0.000
3.174 10.539 8.222 72.563 1.871 3.469
TOASTERS 0 0 0 0 0 0
0.053 0.004 0.266 0.046 0.002 0.000
0.014 0.000 0.265 0.051 0.004 0.000
0.044 0.002 0.511 0.095 0.000 0.000
Table 4: The percentage of discarded pixels averaged over
the length of the animation.
avg %
avg %
avg %
avg %
avg %
avg %
6.0 % 17.9 % 10.7 % 29.9 % 31.6 % 3.6 %
15.3 % 8.3 % 3.1 % 20.3 % 15.1 % 0.8 %
TOASTERS 0.0 % 5.6 % 1.8 % 10.9 % 5.3 % 0.0 %
0.1 % 6.8 % 1.6 % 13.8 % 2.9 % 1.2 %
0.3 % 6.4 % 2.9 % 20.3 % 7.7 % 0.4 %
0.3 % 10.2 % 6.5 % 25.8 % 8.1 % 0.4 %
virtual reality setup, in which the user is constantly
swaying their head slightly during the recording. The
most active compared dataset EMERALDSQUARE has
the highest average in yaw rotation of the previous
work, but it is still 16× smaller than the proposed
dataset. Moreover, its variance is over 5250× smaller
in pitch rotation and 764× smaller in yaw rotation.
In Table 4, we can see that both of the proposed
datasets have a higher percentage of pixels that should
be discarded due to view frustum changes. The aver-
age per-frame frustum discard percentage is 6 15%
with the ETERNALVALLEY datasets, whereas the
highest compared dataset EMERALDSQUARE only
has 0.3%. This lines up with the previously recog-
nized change in the camera’s motion. The same trend
continues with all of the compared properties, ETER-
NALVALLEYFPS having the highest rate of low con-
fidence pixels, and ETERNALVALLEYVR the second
most. The dataset EMERALDSQUARE does have a
reasonably large average percentage with world posi-
tion and shading normal compared to the other ORCA
sets. This is most likely explained due to the amount
of high-frequency vegetation the dataset has, as the
park in the EMERALDSQUARE is filled with bushes
and trees.
The proposed ETERNALVALLEY datasets also
show more low confidence pixels due to lighting con-
ditions: both of them have lot of change in direct
TauBench: Dynamic Benchmark for Graphics Rendering
Figure 3: Path traced frames from different moments in
datasets. On top ETERNALVALLEYFPS, and on the bottom
lighting condition, and ETERNALVALLEYFPS has
the most significant change in the indirect radiance of
all the datasets. BISTROINTERIOR and EMERALD-
SQUARE do also present change in their direct radi-
ance, but less than the proposed sets.
We have presented TauBench: two new datasets,
containing an excellent basis to benchmark tempo-
ral rendering. The datasets present significantly more
temporal complexity than the previously released
ORCA datasets.
The proposed datasets contain actual rendering
settings captured from a game, which more realis-
tically represent the reuse methods’ use of context.
When comparing the camera activity with the posi-
tion and rotational velocity, we show more remark-
able change and variation for the animation duration
than the compared ORCA datasets. We also showed
increased temporal reuse challenge per auxiliary fea-
ture buffer, including world position, shading normal,
direct, and indirect radiance compared to earlier work.
Associated features are often used as input for tempo-
ral rendering methods.
The proposed TauBench datasets are more dy-
namic in many aspects: There are thousands of dy-
namic rigid body objects, whereas the highest com-
pared dataset contains only 15. The TauBench sets
also have 32 skeletal armatures moving around the
scene, with 52 animated bones each, whereas the
compared datasets have none. Moreover, there are
vastly more static and dynamic lights compared to the
previously released rendering datasets. The previous
work mainly relies on the sunlight and the environ-
ment in the dataset, and the highest count of static
lights is in the Bistro Interior with 4 point lights. Both
of the proposed datasets surpass this by having over
70 static point lights and 600 animated point lights
throughout the animation.
In the future, it would be interesting to extend
our dataset comparisons to separate direct and indi-
rect radiance comparisons to diffuse, glossy, trans-
missive, and volumetric, as now they are all com-
pared in a combined sum. Individually handling the
materials would more closely represent how they are
handled in typical renderers. We also invite differ-
ent rendering fields to benchmark their state-of-the-
art reuse methods with these two dynamic datasets,
with perceived and analytical image quality compar-
isons, and the methods compute performance. In ad-
dition to rasterization and path tracing, the dynamic
benchmarks could also be used with real-time virtual
point lights, VR, or light field rendering. It would also
be interesting to have additional comparisons against
other published datasets, like the ray tracing bench-
mark BART (Lext et al., 2001) and the Beeple’s ZE-
RODAY dataset in ORCA (Winkelmann, 2019).
This project has received funding from the ECSEL
Joint Undertaking (JU) under Grant Agreement No
783162 (FitOptiVis). The JU receives support from
the European Union’s Horizon 2020 research and in-
novation programme and Netherlands, Czech Repub-
lic, Finland, Spain, Italy. It was also supported by Eu-
ropean Union’s Horizon 2020 research and innovation
programme under Grant Agreement No 871738 (CP-
SoSaware). The project was also supported in part by
the Academy of Finland under Grant 325530.
Alexa, M., Behr, J., and M
uller, W. (2000). The morph
node. In Proceedings of the fifth symposium on Virtual
reality modeling language (Web3D-VRML).
Amazon Lumberyard (2017). Amazon Lumber-
yard Bistro, Open Research Content Archive
(ORCA). https://developer.nvidia.com/orca/amazon-
Amazon.com, Inc. (2021). Demo files. https://
GRAPP 2022 - 17th International Conference on Computer Graphics Theory and Applications
Benty, N., Yao, K.-H., Clarberg, P., Chen, L., Kallweit,
S., Foley, T., Oakes, M., Lavelle, C., and Wyman,
C. (2020). The Falcor rendering framework. https:
Blender Foundation (2021). Demo files. https://
Bourke, P. (2011). Wavefront .obj file format specification.
Butler, D. J., Wulff, J., Stanley, G. B., and Black, M. J.
(2012). A naturalistic open source movie for optical
flow evaluation. In European Conf. on Computer Vi-
sion (ECCV), Part IV, LNCS 7577. Springer-Verlag.
Gilabert, M. and Stefanov, N. (2012). Deferred illumi-
nation in Far Cry 3. https://www.gdcvault.com/
Gribble, C. P., Ize, T., Kensler, A., Wald, I., and Parker,
S. G. (2007). A coherent grid traversal approach
to visualizing particle-based simulation data. IEEE
Transactions on Visualization and Computer Graph-
ics, 13(4).
Haines, E. (1987). Neutral File Format. http://
Hanika, J., Dammertz, H., and Lensch, H. (2011). Edge-
a-trous wavelets for local contrast enhance-
ment with robust denoising. In Computer Graphics
Forum, volume 30. Wiley Online Library.
Hedman, P., Karras, T., and Lehtinen, J. (2017). Sequential
Monte Carlo instant radiosity. IEEE Transactions on
Visualization and Computer Graphics, 23(5).
Kalantari, N. K., Bako, S., and Sen, P. (2015). A machine
learning approach for filtering Monte Carlo noise.
ACM Transactions on Graphics, 34(4).
Keller, A. (1997). Instant radiosity. In Proceedings of the
24th Annual Conference on Computer Graphics and
Interactive Techniques.
anek, J., Gautron, P., Pattanaik, S., and Bouatouch, K.
(2005). Radiance caching for efficient global illumi-
nation computation. IEEE Transactions on Visualiza-
tion and Computer Graphics, 11(5).
Lext, J., Assarsson, U., and Moller, T. (2001). A benchmark
for animated ray tracing. IEEE Computer Graphics
and Applications, 21(2).
Li, T.-M., Wu, Y.-T., and Chuang, Y.-Y. (2012). SURE-
based optimization for adaptive sampling and recon-
struction. ACM Transactions on Graphics, 31(6).
Majercik, Z., Guertin, J.-P., Nowrouzezahrai, D., and
McGuire, M. (2019). Dynamic diffuse global illu-
mination with ray-traced irradiance fields. Journal of
Computer Graphics Techniques, 8(2).
McCool, M. D. (1999). Anisotropic diffusion for Monte
Carlo noise reduction. ACM Transactions on Graph-
ics, 18(2).
Mueller, J. H., Neff, T., Voglreiter, P., Steinberger, M., and
Schmalstieg, D. (2021). Temporally adaptive shading
reuse for real-time rendering and virtual reality. ACM
Transactions on Graphics, 40(2).
Nicholas Hull, K. A. and Benty, N. (2017). NVIDIA
Emerald Square, Open Research Content Archive
(ORCA). https://developer.nvidia.com/orca/nvidia-
Oortmerssen, W. v. (2021). Cube 2: Sauerbraten. http:
Ramey, D., Rose, L., and Tyerman, L. (1995). MTL mate-
rial format (Lightwave, OBJ). http://paulbourke.net/
Scherzer, D., Yang, L., Mattausch, O., Nehab, D., Sander,
P. V., Wimmer, M., and Eisemann, E. (2012). Tem-
poral coherence methods in real-time rendering. In
Computer Graphics Forum, volume 31. Wiley Online
Sen, P. and Darabi, S. (2012). On filtering the noise from the
random parameters in Monte Carlo rendering. ACM
Transactions on Graphics, 31(3).
Sung-eui, Y. (2014). Korea Advanced Institute of Science
and Technology (KAIST) model benchmarks. https:
Tamstorf, R. and Pritchett, H. (2019). The challenges of
releasing the Moana Island Scene. In Eurograph-
ics Symposium on Rendering - DL-only and Industry
Track. The Eurographics Association.
Tatzgern, W., Mayr, B., Kerbl, B., and Steinberger, M.
(2020). Stochastic substitute trees for real-time global
illumination. In Symposium on Interactive 3D Graph-
ics and Games.
Tawara, T., Myszkowski, K., and Seidel, H.-P. (2004). Ex-
ploiting temporal coherence in final gathering for dy-
namic scenes. In Proceedings Computer Graphics In-
ternational, 2004. IEEE.
The GAMMA research group at University of North Caro-
line (2018). UNC dynamic scene benchmarks. https:
Wald, I. (2019). Utah animation repository. https://
Wald, I., Benthin, C., and Slusallek, P. (2003). Interactive
global illumination in complex and highly occluded
environments. In Rendering Techniques.
Wald, I., Boulos, S., and Shirley, P. (2007). Ray tracing
deformable scenes using dynamic bounding volume
hierarchies. ACM Transactions on Graphics, 26(1).
Walt Disney Animation Studios (2018). Moana Is-
land Scene (v1.1). https://www.disneyanimation.com/
Ward, G. J. and Heckbert, P. S. (1992). Irradiance gradi-
ents. Technical report, Lawrence Berkeley Lab., CA
(United States); Ecole Polytechnique Federale.
Winkelmann, M. (2019). Zero-Day, Open Research Content
Archive (ORCA). https://developer.nvidia.com/orca/
Yang, L., Liu, S., and Salvi, M. (2020). A survey of tem-
poral antialiasing techniques. In Computer Graphics
Forum, volume 39. Wiley Online Library.
Yang, L., Nehab, D., Sander, P., Sitthi-amorn, P., Lawrence,
J., and Hoppe, H. (2009). Amortized supersampling.
ACM Transactions on Graphics, 28(5).
TauBench: Dynamic Benchmark for Graphics Rendering