VCrowds.pdf

(496 KB) Pobierz
LOD Human Representations: A Comparative Study
Rachel McDonnell, Simon Dobbyn, Carol O’Sullivan
Image Synthesis Group
Trinity College Dublin, Dublin 2, Ireland
Telephone: +353 1 608 1220, Fax: +353 1 677 2204
E-mail: Rachel.McDonnell@cs.tcd.ie, Simon.Dobbyn@cs.tcd.ie, Carol.OSullivan@cs.tcd.ie
Figure 1:
(a) crowd system (b) participant taking part in the experiments
Abstract
Real-time crowd systems are increasing in popularity and importance. With this comes the need for accurate low
level of detail (LOD) human model representations. Currently, the two most commonly used representations for
large scale crowds are image based representations and low level geometric meshes. In this paper, we evaluate
the effectiveness of the two representations in terms of their appearance and how accurately they convey motion.
We show that low resolution geometry is not equivalent to image based representations at replicating the motion
of high resolution geometry. We also report the distances from the camera at which impostors and different
simplified meshes are perceptually equivalent to high resolution geometry.
Keywords
Perception, evaluation, LOD, impostors.
1. Introduction
As the sophistication of real-time applications increases, so does the need for realism.
Environments like cities are populated by crowds in the real world, but are often left empty in
computer simulations such as games or entertainment. For this reason, real-time crowd
systems are becoming increasingly important in the graphics community. Much recent
research has focused on the most computationally efficient ways to render real-time crowds in
order to populate these environments. The focus has been on investigating the rendering costs
of the numerous ways of graphically modeling and animating virtual humans. Ideally, crowds
of high resolution polygon meshes would be used to populate scenes, but rendering costs
forbid large numbers of these meshes to be displayed in real-time. For this reason, developers
make use of lower LOD representations. Image based representations or “impostors” and low
resolution polygon meshes are the two most commonly used substitutes at present. In
previous work [1], we showed that impostors are an effective substitute for high resolution
meshes when displayed at certain distances from the camera. We now compare the relative
effectiveness of the low resolution mesh. We address the questions: Are low resolution
meshes and impostors equally effective at replicating the motion of a high resolution mesh?
How much can you decrease the resolution of a mesh by, while still having it perceptually
equivalent to the high resolution mesh, and at what distance from the camera? How does this
compare to the distance at which you can notice an impostor and is it as cost effective in
terms of rendering?
In our research, we use the scientific discipline of Psychophysics to explore the relationship
between physical stimuli and subjective responses so that we can investigate the effect of
model representation on perception with regards to motion and appearance. We present a
study that analyzes people’s ability to perceive variation in motion when using different
representations. We also perform a series of perceptual experiments to investigate human
sensitivity to different model representations based on their appearance.
2. Background
Previous work on real-time crowd systems was focused on developing the most efficient ways
of displaying large crowds. Tecchia et al. [2] populated an environment with thousands of
animated characters at run-time using pre-generated impostors. Aubel et al. [3] showed how
dynamic impostors can be used to render virtual humans. De Heras Ciechomski et al. [4]
designed a real-time virtual audience for a virtual heritage reconstruction. They chose low
level geometry meshes rather than impostors to populate their environment, as they felt that
the viewer would be too close to the virtual audience to use impostors, who can look
pixillated at close distances. In [5] we presented a LOD rendering system for crowds, where
seamless interchanges between pre-generated impostors and geometry model representations
were implemented in order to balance visual quality and performance (Figure 1a).
Previous work on perception of human motion in the context of computer graphics has mainly
been focused on the effect of animation quality on user perception. Wang et al. [6] conducted
a set of experiments to evaluate a cost function proposed by Lee et al. [7] for determining the
transition quality between motion clips. Other recent work by Harrison et al. [8] examined the
perceptual impact of dynamic anomalies in human animation. Reitsma and Pollard [9]
conducted a study, developing a metric to evaluate the perceived error introduced during
motion editing. Harrison et al. [10] focused on higher-level techniques for specifying and
modifying human motions. Oesker et al. [11] investigated the extent to which observers
perceptually process the LOD in naturalistic character animation. The study most related to
our work is by Hodgins et al. [12]. They performed a series of perceptual experiments, the
results of which indicated that a viewer’s perception of motion characteristics is affected by
the geometric model used for rendering. Participants were shown a series of paired motion
sequences and asked if the two motions in each pair were the “same” or “different”. The
motion sequences in each pair were rendered using the same geometric model. For the three
types of motion variation tested, sensitivity scores indicated that subjects were better able to
observe changes when viewing the polygonal model than they were with a stick figure model.
While there has been extensive research on the numerous ways of graphically representing
virtual humans (including their associated rendering costs), there has been little research
conducted on perceptually evaluating these representations. In Hamill et al. [1], we evaluated
the effectiveness of impostor representations for real-time rendering and animation of static
buildings and dynamic virtual humans, and investigated their impact on the perception of
human motion. We found that impostors were an effective substitute for detailed geometry in
the targeted application area (i.e., large-scale simulations inhabited by crowds).
In this paper, we extend the virtual human representation evaluation of our previous study
with a broader study of different LODs for crowds in order to improve the realism of our
crowd system, while maintaining interactive frame rates. We also aim to establish an effective
experimental procedure to evaluate human representations that can be used by developers of
other crowd systems.
We used Psychophysics to explore the connection between physical stimuli and subjective
responses [13]. In a typical psychophysical experiment, subjects are asked to compare various
stimuli to a reference stimulus and report if they thought that they were the “same” or
“different”. A statistical curve called a Psychometric Function is fitted to the cumulative
responses of the experiment to provide a way of analysis. A simple Ogive inverse normal
distribution can be fitted to evenly distributed stimulus data, while a logistical function such
as that described by Linschoten et al. [14] can be fitted to experimental data where the
probability of detection of a stimulus falls below a chance detection level.
The Point of Subjective Equality (PSE) and the Just Noticeable Difference (JND) are two
typical values that can be estimated from the fitted curve. The PSE is the threshold at which
participants are able to tell two stimuli apart. This is typically estimated as the 50% detection
level on the best fit psychometric curve (i.e., the point where same and different responses are
equally likely). The JND is the smallest difference in intensity required for a person to
distinguish 2 stimuli. The JND is typically estimated as the amount of additional stimulus
needed to increase a participant’s detection rate from 50% to 75% on the fitted psychometric
function. It is a measure of the participant’s uncertainty; the larger the value of the JND, the
more uncertain they were in making their choices. The steeper the function, the smaller the
JND. ANalysis of Variance (ANOVA) is used to test the null hypothesis that two means are
equal. The null hypothesis is rejected if there are significant differences between the means.
3. Perception of Human Motion
In a LOD crowd system that simultaneously displays different model representations, as
described in [5], it is important that the quality of the motion of the lower LODs is not
significantly different from that of the high resolution. As discussed in Section 2, Hodgins et
al. [12] showed that model type affected user perception of human motion, when a stick
figure model’s motion was compared to a polygonal model. We found in Hamill et al. [1] that
the motion of the impostor accurately replicated the motion of the high resolution model. We
now test whether or not the low resolution polygon mesh replicates the motion of the high
resolution mesh as accurately as the impostor, using the same psychophysical procedure. We
also test the performance of a stick figure model to compare our results to those of Hodgins et
al. [12] and a point light source model as a baseline test.
3.1.
Model Types
Five different representations of a male model were used (Figure 2). Two of the models were
polygonal models with deformable meshes which were manipulated by an underlying
skeleton; the high resolution polygon model had a deformable mesh of 2022 polygons, while
the low resolution polygon model had only 808 polygons for a deformable mesh. The low
resolution model was created by applying the 3D Studio Max
multires
modifier to the high
resolution model. The modifier allows one to manually maintain part of the mesh at full
resolution while reducing the LOD of the rest. Initially, we chose this option in order to keep
a high number of polygons around the areas that would be deformed most by the joints.
However, this manual selection is only necessary when simplifying meshes to very low
resolutions, so we used automatic simplification. We automatically simplified the mesh as
much as possible, without making the simplified version look different from the original,
resulting in a mesh of 40% of the number of vertices of the original.
Figure 2:
High resolution, low resolution, impostor, stick figure, point light source model
Impostors were the third type used and were created by mapping an image of the 3D object
onto a quadrilateral and orientating this quadrilateral towards the viewer. As in Hamill et al.
[1], a pre-generated impostor approach was used, which involved the offline rendering of the
human model’s high resolution polygonal model from 17*8 different viewpoints around the
model for 10 animation keyframes. All geometric and impostor representations were
dynamically lit in the experiments. A stick figure was the next type and was created by
drawing lines between the joints of the underlying skeleton. This representation was used in
order to compare our findings with those of Hodgins et al. [12]. Studies have shown that 13
moving light points, attached to the joints of the body, suffice to create a rich perception of a
moving human figure [15]. Using only 13 dots to display a human is the simplest
representation and it is also the least computationally expensive of the 5 models, so we
included this representation as the lowest LOD.
3.2.
Method and Participants
We used a between-groups design for this experiment, where each group viewed a different
model representation. This approach was chosen as we felt that, if allowed to view all of the
models, the participants might base their judgments on characteristics of the models rather
than the actual differences in motion, e.g., the impostor images contain artefacts along the
edges which may cause the participant to focus on the artefacts instead of the overall motion
if they had already seen the geometric model which showed less artefacts.
Sixty-five participants (23 females, 42 males, aged between 17 and 35) took part in the
experiment and were given book tokens as a reward for participation. They ranged from
college staff, postgraduate students, undergraduate students and professionals and were all
from different educational backgrounds. All participants were naive as to the purpose of the
experiment and had normal or corrected to normal vision. The experiment was displayed on a
21 inch flat screen C.R.T. monitor. A grey-scale checkerboard floor plane was used so that
the movement of the model could be seen clearly. All models were rendered in grey-scale in
order to eliminate any bias due to colour that may have occurred. Lighting and rendering
conditions were constant throughout the experiment.
3.3.
Creating the Motion Variation
A reference motion
R
was created which consisted of 10 frames of a key-framed walk
motion. This motion was cyclic and was repeatedly looped until 4 seconds of animation were
recorded. The 10 frames of
R
were modified a number of times to create the arm, leg, and
torso motion variation sequences. The arm and the torso variations were chosen as they were
also used by Hodgins et al. [12].
Figure 3:
10 frames of animation of the reference motion
R.
kA1
is the frame highlighted on the left,
kA2
is the
frame highlighted on the right.
Firstly, the performance of the participants in distinguishing smaller and larger dynamic arm
motions was examined. Assessing the arm motion variation involved comparing
R
to a set of
motions which altered the distance of the arm from the body at certain keyframes.
kA1
and
kA2
were the keyframes in
R
where the arms were furthest away from the body (Figure 3).
kA1
and
kA2
were modified by a fixed amount 10 times, and the resulting 10 motions
represented the 10 different steps in the staircase analysis. The modifications were made by
rotating the upper left arm joint in
kAl
at the shoulder along the positive horizontal axis by a
fixed number of degrees. The right arm was altered by the same amount in the reverse
direction. The pose of the skeleton at
kA1
was then copied and the inverse pose was pasted
onto the skeleton at
kA2.
The 10 altered biped motion sequences were then exported and loaded into an OpenGL
rendering system and applied to our high and low resolution models with deformable meshes.
The stick figure and point light source models were also rendered using these motion
sequences. All of the altered motions were cyclic and looped until 4-second movies could be
recorded. Forty of these movies were recorded (10 for each model type). The 10 impostor
sequences were then rendered from the high resolution polygon model and recorded as
movies.
A similar test was conducted to test the ability of the participants in distinguishing larger and
smaller leg motions for all representations. A further set of 50 motion sequence movies was
created in a similar manner to the arm motions, except that the leg was altered by iterative
translations along the longitudinal and vertical axes.
Finally, the ability of the participants to distinguish modifications to the torso was tested. A
further 50 movies were created by making kinematic alterations to
R.
In this instance, the
alterations were made by iteratively rotating the lower spine of the skeleton by a fixed number
of degrees around the longitudinal axis.
3.4.
Experiment Procedure
Participants viewed pairs of movies, and were asked to specify whether they thought that the
motion of the character in the movies was the “same” or “different” (Figure 1b). They were
told to judge difference based on the overall motion of the character and not to focus on
colour, speed or any other factors which were constant throughout the experiment. After the
first 4-second movie was viewed, the participant pressed a “view next” button on the screen
using the mouse. The next movie was then presented for 4 seconds and the participant had to

Plik z chomika:

Inne pliki z tego folderu:

Inne foldery tego chomika:

Zgłoś jeśli naruszono regulamin