A PSYCHOPHYSICAL STUDY OF FOVEAL GRADIENT BASED

SELECTIVE RENDERING

Veronica Sundstedt

Department of Computer Science, Bristol University, Bristol, United Kingdom

Keywords:

Perception, user studies, attention, selective rendering, inattentional blindness, importance maps.

Abstract:

High-ﬁdelity rendering of complex scenes at interactive rates is one of the primary goals of computer graph-

ics. Since high-ﬁdelity rendering is computationally expensive, perceptual strategies such as visual attention

have been explored to achieve this goal. Inattentional Blindness (IB) experiments have shown that observers

conducting a task can fail to see an object, although it is located within the foveal region (2

◦

). However,

previous attention based algorithms assumed that IB would be restricted to the area outside the foveal region,

selectively rendering the areas around task-related objects in high quality and the surrounding areas in lower

quality. This paper describes a psychophysical forced-choice preference experiment assessing if participants,

performing a task or free-viewing animations, would fail to notice rendering quality degradation within the

foveal region. The effect of prior knowledge on the level of perceived quality is also studied. The study in-

volves 64 participants in four conditions: performing a task, or free-viewing a scene, while being naive or

informed about assessing rendering quality. Our results show that participants fail to notice the additional

reduction in quality, decreasing the overall computation 13 times. There was also a signiﬁcant difference in

the results if free-viewing participants were informed.

1 INTRODUCTION

Obtaining realistic images has always been one of the

major goals in computer graphics. Applications rang-

ing from entertainment, lighting design, to archaeo-

logical reconstructions, and scientiﬁc visualisations

have required realistic models of light propagation

and scattering. One traditional method for speeding

up global illumination calculations is to distribute the

workload over several machines, each processing a

part of the scene in parallel. However, even in the

parallel case, rendering efforts are spent on improv-

ing details that would not be perceived by a human

observer.

For many applications, rather than using more ma-

chines to accelerate the rendering, it is more efﬁcient

to reduce the number of computations needed. It is

attractive to improve the efﬁciency of rendering by

performing less work. Hence perceptually-based ren-

dering algorithms have become an important research

topic in computer graphics. The goal of perceptually-

based rendering algorithms is to signiﬁcantly reduce

computation that is necessary to obtain an image that

is perceptually indistinguishable from a fully con-

verged solution or gold standard, which is a hypo-

thetical perfect rendering (Woolley et al., 2003). Re-

cently, models of the human visual system (HVS), in

particular those based on visual attention processes,

have been used in perceptually assisted renderers to

make progress towards this goal.

1.1 Visual Attention Processes

Human visual perception is a selective process in

which a part of the observed environment is chosen

for further processing in the visual cortex of the brain.

Although, the perception of an environment does not

only depend on the sensory input of the observer, but

also on the visual task performed (Cater et al., 2003).

When free viewing a scene, the attentional processes

of the observer are guided by automatic low-level vi-

sion. These are normally referred to as bottom-up pro-

cesses (James, 1957). Low-level, bottom-up features

which inﬂuence visual attention include contrast, size,

shape, colour, brightness, orientation, edges, and mo-

tion (Itti et al., 1998). In contrast, when perform-

ing a visual task it is directed in the pursuit of our

goals (Yarbus, 1967). These are normally referred to

207

Sundstedt V. (2008).

A PSYCHOPHYSICAL STUDY OF FOVEAL GRADIENT BASED SELECTIVE RENDERING.

In Proceedings of the Third International Conference on Computer Graphics Theory and Applications, pages 207-214

DOI: 10.5220/0001097402070214

 SciTePress

(a) (b) (c) (d)

Figure 1: Map examples from a corridor scene (Frame 1): (a) high quality rendering, (b) saliency map, (c) task objects, and

(d) task map with foveal gradient angle.

as top-down processes (James, 1957). In both cases,

the HVS focuses its attention on certain objects at the

expense of other details in the scene. The inatten-

tional blindness (IB) phenomenon (Mack and Rock,

1998) relates to our inability to perceive features or

objects in a visual scene if we are not attending to

them. It is suggested that IB occurs because the

observer is focussed on performing a speciﬁc task,

i.e. there is no conscious perception without atten-

tion (Mack and Rock, 1998).

1.2 Visual Attention Models

One method of reducing computation, while main-

taining a result with high perceptual quality, is to

adapt the rendering parameters of the image based on

models of human visual attention processes. In this

way areas which are perceptually more relevant will

receive further improvement. This results in an image

with a spatially shifting degree of accuracy, referred

to as selective rendering. In selective rendering algo-

rithms the bottom-up process is often modelled using

a saliency map derived from the computational model

developed in (Itti et al., 1998). This model extract

salient features based on colour, intensity, and orien-

tation. An example of a saliency map for the corridor

scene is shown in Figure 1 (b).

The idea of using task maps to model the top-

down process in selective rendering was introduced

in (Cater et al., 2003). A task map is a grey-scale im-

age, which consists of objects related to the task in

white and the surrounding areas in black, as shown

in Figure 1 (c). There is also an option to gradu-

ate the shade between black and white over an area

that corresponds to the image registered on the foveal

region (2

◦

), mimicking the high visual acuity of the

HVS (Snowden et al., 2006), as shown in Figure 1

(d).

Previous rendering algorithms using task maps

have assumed that IB would be restricted to the area

outside the foveal region. However, in (Mack and

Rock, 1998) it was shown that participants failed to

notice a critical stimulus, even though it appeared

within the centre of their ﬁeld of view and coincided

with a ﬁxation. This inspired the psychophysical

forced-choice preference experiment presented in this

paper. The psychophysical experiment investigates

if previous work can be improved by using selective

rendering within the foveal region in the presence of a

high-level task focus or when free-viewing a scene. It

also studies if the effect of prior knowledge (being in-

formed about assessing rendering quality) would alter

the level of perceived quality.

2 RELATED WORK

Extensive overviews of perceptually adaptive graph-

ics techniques are given in (McNamara, 2001;

O’Sullivan et al., 2004). A more recent and com-

prehensive summary of different selective rendering

techniques are presented in (Debattista, 2006; Sund-

stedt, 2007). The following sections describe render-

ing techniques which have taken into account visual

attention models.

2.1 Gaze-Contingent Techniques

Gaze-contingent displays (GCDs) (Loschky et al.,

2003) track the user’s attention using an eye-tracker

and render information in full detail only at the ob-

server’s current focus of attention. To prevent ob-

servers from noticing the lower quality in the periph-

eral regions, the size of the foveal regions is based

on the the extent of the user’s perceptual span (2

◦

Focus plus context screens include both high and low

detail by combining a wall-sized low-resolution dis-

play with an embedded high-resolution screen (Baud-

isch et al., 2003). The user uses the mouse to pan

the display content into the high-resolution area. Eas-

ily perceived displays aim to direct the attention of

GRAPP 2008 - International Conference on Computer Graphics Theory and Applications

208

the viewer (Baudisch et al., 2003). This has been ex-

ploited in for example art where eye-tracking infor-

mation from one participant has been used to create

aesthetically pleasing images.

2.2 Level of Detail

A bottom-up attention model for Level of Detail

(LOD) simpliﬁcation taking into account size, posi-

tion, motion, and luminance was presented in (Brown

et al., 2003). The mesh saliency algorithm in (Lee

et al., 2005) exploits the centre-surround operation

from (Itti et al., 1998) to generate more visually pleas-

ing images by prioritising the geometry as the most

important salient feature. In (Howlett et al., 2005)

polygon models were simpliﬁed based on saliency

identiﬁed using eye-tracking of human observers.

In (Yang and Chalmers, 2005) the LOD was reduced

of objects not related to the task being performed by

the observer. A perceptual stategy for collision detec-

tion was proposed in (O’Sullivan, 2005). For exam-

ple, a lower LOD could be used between objects that

are not being focused upon.

2.3 Attention based Rendering

A ﬁrst attempt to account for selective visual atten-

tion in global illumination rendering of dynamic en-

vironments was proposed in (Yee et al., 2001). Here

the saliency model (Itti et al., 1998) was extended to

include motion. An Aleph map was created that com-

bined the saliency map with a spatiotemporal contrast

sensitivity function (CSF) (Myszkowski et al., 1999).

The Aleph map was used to guide the search radius

accuracy for the interpolation of irradiance cache val-

ues so that perceptually important regions would be

more accurate. This made the indirect lighting com-

putations more efﬁcient since larger error can be tol-

erated in less salient regions.

One of the ﬁrst uses of an attention model in a

real-time application was proposed in (Haber et al.,

2001). One of the problems with interactive walk-

throughs is to render non-diffuse objects in real-

time. The non-directional light component can be

efﬁciently pre-computed, whereas directional light

from glossy surfaces, for example, can only be com-

puted during execution. Haber et al. rendered a pre-

computed global illumination solution using graphics

hardware while updating non-diffuse objects dynam-

ically with a ray tracer. The selection of these types

of objects was based on the saliency model (Itti et al.,

1998). Haber et al. also took top-down behaviour into

account by weighting the saliency model with bias to-

wards objects in the centre.

The idea of exploiting IB in a rendering frame-

work was ﬁrst proposed in (Cater et al., 2002). Cater

et al. investigated if areas in a scene that normally

would attract attention would do so in the presence of

a task. In their experimentsobservers were shown two

animations. One was rendered in high quality and the

other one either in low quality or selectively compos-

ited. The selectively composited animation had only

the pixels in the 2

◦

foveal region centred around the

location of the task object in high quality. The qual-

ity was then blended to lower quality within an angle

of 4

◦

based on ﬁndings by (McConkie and Loschky,

1997).

Half the subjects were instructed to simply watch

the animations and the other half were asked to per-

form a task. After completion of the experiment the

participants were asked if they noticed a quality dif-

ference. The results showed a signiﬁcant difference

for all pair-wise comparisons to high quality while

free-viewing the animations. While performing a task

there was only a signiﬁcant difference between the

high quality and low quality comparison. Cater et

al. (Cater et al., 2003) used these results as a basis for

a task-based perceptual rendering framework, which

combined predetermined task maps with a spatiotem-

poral CSF to guide a progressive animation system.

3 SELECTIVE RENDERING

To selectively render the stimuli used for the ex-

periment presented in this paper a region-of-interest

(ROI) rendering system was used (Debattista, 2006;

Sundstedt, 2007). For the convenience of the reader,

the technical details that underpins this work is brieﬂy

reviewed. The rendering system is composed of two

major processes:

Region-of-interest (ROI) guidance uses a combi-

nation of saliency and a measure of task relevance

to direct the rendering computation (in a com-

bined importance map).

Selective rendering corresponds to the traditional

rendering computation (Ward, 1994). However,

computational resources are focused on parts of

the image which are deemed more important by

the ROI guidance.

The process begins with a rapid image estimate

(in the order of ms) of the scene using a quick ras-

terisation pass in hardware (Longhurst, 2005). This

estimate can be used in two ways. Firstly for building

the task map by identifying user-selected task objects,

and secondly, by using it as an input to a saliency gen-

erator. During creation of the task map the program

A PSYCHOPHYSICAL STUDY OF FOVEAL GRADIENT BASED SELECTIVE RENDERING

209

Observer’s eye

Screen

Figure 2: Relation between visual angle and pixel radius

on screen.

reads in the geometry information and a list of prede-

ﬁned task objects. It then produces a map with task

objects in white and the other geometry in black, as

shown in Figure 1 (c). The task map can also take

into account the area the fovea in the eye covers in the

environment. The foveal region in the eye, where the

visual acuity is highest, is only the central 2

◦

of the

visual ﬁeld.

When an observer is watching the environment,

this area corresponds to a region on an object. In the

context of this paper this object is a computer monitor.

Although the size of the foveal region is ﬁxed within

the eye, the area it covers on the object varies with the

distance (d) between the eye and the object. As the

size of regions that project an image onto the foveal

region changes, for simplicity the distance on the ob-

ject as an angle (A) subtended at the eye is measured,

as shown in Figure 2. The radius (R) of this circular

region can be measured in pixels if the ratio between

the screen resolution and monitor size is known, as

computed by equation 1:

R = ratio(d· tanA) (1)

The region on the screen is a circle with radius 1

◦

but for simplicity squares with a width and height of

◦

are used. To account for a gradient (which blends

the quality from high to low), a graduated ﬁll can be

used between a square of 2

◦

and 4

◦

, as shown in Fig-

ure 1 (d). This graduated ﬁll was proposed in (Cater

et al., 2002) after ﬁndings described in (McConkie

and Loschky, 1997).

In the creation of the saliency map the image esti-

mate serves to locate areas where an observer will be

most likely to look. The estimate of the scene con-

tains only direct lighting, but a simple shadow and

reﬂection calculation is also included. The saliency

estimation is carried out by using the existing method

proposed in (Itti et al., 1998) and is computed in 2-3

seconds per frame.

A hardware implementation can generate a

saliency map in the order of tens of millisec-

onds (Longhurst, 2005). The two maps have pre-

viously been used separately and in combination to

form an importance map (IM) which accounts for

both the bottom-up and top-down processes (Sund-

stedt et al., 2005). The values in the importance map

are used to direct the rendering. However, this paper

is only using the task map.

3.1 Previous Visual Trial

In (Sundstedt et al., 2005) a psychophysical experi-

ment was performed which showed that both saliency

maps and task maps (including the foveal region gra-

dient) can be used successfully to selectively render

in high quality only the important areas of animations

while performing a task or free-viewing a scene. The

experiment investigated how models of visual atten-

tion, low-level and task-dependent on their own, and

as a hybrid wouldwork in a selective rendering frame-

work. 160 participants took part in this trial (124 men

and 36 women; age range: 18-39). There were in total

ten groups with 16 participants in each group.

Each participant was shown a high quality (HQ)

animation and a HQ, low quality (LQ) or selectively

rendered animation of the corridor scene, shown in

Figure 1 (a). The selectively rendered animations

were either generated using a task map (TQ), saliency

map (SQ) or a linear combination of the two (IQ).

Two HQ animations were shown in one group to con-

clude if where performance was caused by chance,

the chance performance was not based on there be-

ing a strong preference for one of the two scenes

(p > 0.05). The HQ animations were rendered us-

ing an extended version if the Radiance (Ward, 1994)

renderer

rpict

using 16 rays per pixel and a specular

threshold value of 0.01. The LQ animations were ren-

dered using 1 ray per pixel and a specular threshold

value of 1. The extended rendering system is further

described in (Debattista, 2006; Sundstedt, 2007).

Half the subjects were asked to perform a task,

in this case counting the items related to ﬁre safety,

whereas the other group were only shown two anima-

tions without any previous instructions. After com-

pletion of the experiment, each participant determined

which of the two animations they thought had the

worse rendering quality. The results conﬁrmed, us-

ing a one-sample chi-square test (df=1, critical value

3.84 at 0.05 level of signiﬁcance), that for both par-

ticipants performing a task and free-viewing in the

HQ/TQ, HQ/SQ and HQ/IQ, the difference in propor-

tions was not signiﬁcant (p > 0.05). From this it was

concluded that the two animations were perceived as

GRAPP 2008 - International Conference on Computer Graphics Theory and Applications

210

a similar quality. For participants free-viewing and

performing a task in the HQ/LQ condition there was

a signiﬁcant difference (p < 0.05).

As IB is described in (Mack and Rock, 1998), ob-

servers conducting a task can fail to see an object, al-

though it is located within the foveal region. Based on

this fact and the result of the previous visual trial an

additional experiment was constructed. The new psy-

chophysical experiment presented in this paper inves-

tigates whether participants fail to notice a reduction

in quality within the foveal region in the presence of a

high-level task focus or when free-viewing a scene. It

also studies if the effect of prior knowledge (being in-

formed about assessing rendering quality) would alter

the level of perceived quality.

4 NO FOVEAL REGION STUDY

Based on the result in (Mack and Rock, 1998) it is hy-

pothesised that in the presence of a high-level task fo-

cus participants would fail to notice quality degrada-

tions even within the foveal region. The relationship

between the foveal region and an area on the screen is

previously explained. This region might contain other

objects unrelated to the task.

If this hypothesis is true, non-task areas within the

foveal region can be rendered with lower quality than

the task objects saving computation. To study if par-

ticipants would fail to notice the reduction in quality

a modiﬁed task map was used for the corridor scene

animations. This time the fovea angle gradient was

excluded, as shown in Figure 1 (c).

The categorical variables in the experiment are:

condition (HQ/TWFQ - Task Without Fovea Quality)

and preference (correct/incorrect). The condition is

the manipulated independent variable and the prefer-

ence is the dependent variable. The two hypotheses

tested in the experiment are stated below:

Hypothesis 1: There is no signiﬁcant difference in

the level of correct/incorrect responses for partic-

ipants performing a task in or free-viewing an an-

imation rendered in HQ and one rendered selec-

tively with only the task objects in HQ (TWFQ).

Hypothesis 2: There is no signiﬁcant difference in

the level of correct/incorrect responses between a

participant being naive to the purpose of the ex-

periment and informed participants.

4.1 Participants and Setup

64 participants took part in the experiment (51 men

and 13 women; age range: 20-37). The participants

taking part in the experiments were all undergraduate

or graduate students. Subjects had a variety of expe-

rience with computer graphics, and all self-reported

normal or corrected-to-normal vision. There were in

total four groups with 16 participants in each group.

Two groups were performing a task, the other two

were free-viewing the scene. To study the effect of

prior knowledge half the participants were naive, or

uninformed, to the forced-choice preference. The

other half were aware that they would be asked about

the quality of the animations they had seen. They

were thus informed as to the purpose of the experi-

ment.

All stimuli were presented on a 17” LCD moni-

tor (1280× 1024 resolution, 60 Hz refresh frequency).

The effect of ambient light was minimised. The par-

ticipants were seated on an adjustable chair, with their

eye-level approximately level with the centre of the

screen. The viewing distance from the participants

to the screen was around 60 cm. All stimuli were

rendered at 900× 900 resolution and displayed in the

centre of the screen with a black background.

4.2 Stimuli

Two different walkthroughs of a corridor scene were

used, which are termed corridor A, and corridor B.

This was done to avoid familiarity effects that might

have inﬂuenced the scan path of the observers. Both

the animations were rendered with different views

and the location of the objects changed within each

scene. Each animation contained the same number of

task-related objects (15) but not the same number of

non-task related objects. An identical type of camera

path was used for both animations.

For both corridor scenes a new animation was

rendered, without the foveal angle gradient added.

This stimuli was termed TWFQ. Using the TWFQ

maps only the task objects were rendered in HQ

and the remaining regions in LQ. The same HQ

animations were used as in the previous visual trial.

Figure 3 shows the timing comparison between a

HQ, TQ, LQ animation used in the previous visual

trial, and a new TWFQ animation. Rendering the new

TWFQ frames took on average 10 minutes, which

is around four times faster then the TQ renderings.

Rendering the entire frame to the same detail as the

task objects in the new TWFQ renderings therefore

took on average 13 times longer. Computing the

TWFQ condition was cheaper than all other selec-

tively rendered stimuli presented in the previous

visual trial (Sundstedt et al., 2005).

A PSYCHOPHYSICAL STUDY OF FOVEAL GRADIENT BASED SELECTIVE RENDERING

211

100

150

200

0 50 100 150 200 250 300

Time (minutes)

Frame No.

TWFQ

Figure 3: Timing comparison for the corridor scene be-

tween a high quality (HQ), foveal region quality (TQ), low

quality (LQ) animation used in the previous visual trial, and

a new no foveal region (TWFQ) animation.

4.3 Procedure

The same experimental setup and method as in the

previous visual trial was used for consistency. The

animations were displayed in the centre of the screen

with a black background. Each animation was 17

seconds long, including a countdown before the

animations started. Following a verbal introduction

to the experiment, each participant was shown two

animations, one from corridor A and one from corri-

dor B. One of the animations was always HQ while

the other one was a selectively rendered animation

using the new TWFQ animation. Each animation was

viewed only once. The order in which the subjects

saw their two animations was also altered to avoid

any bias.

Before beginning the experiment, half the sub-

jects read a sheet of instructions on the procedure

of the particular task they were to perform. As

before, these subjects were asked to take the role

of a ﬁre security ofﬁcer whereby the task was to

count the total number of ﬁre safety items in each

of the two animations. Each of the participants in

these groups were shown an example of what kind

of ﬁre emergency items the scene could contain. For

the participants performing a task while watching

their stimuli, it was also conﬁrmed that the task was

understood prior to the start of the experiment. The

participants in the other half were simply shown the

animations. After completion of the experiment, both

groups were asked which of the two animations they

thought was of worse quality. If a participant could

not determine which one they thought had the worse

quality, they were asked to choose either A or B

(2AFC).

4.4 Results

Figure 4 shows the results of the experiment. In each

pair of conditions, a result of 50% (8 out of 16 partic-

ipants) correct selection in each case is the unbiased

ideal. This is the statistically expected result in the

absence of a preference towards one scene, and indi-

cates that no differences between the high quality and

a lower quality animation were perceived. The results

for the participants performing a task are shown to

the left and the free-viewing results are shown to the

right. The left bar in each graph shows the result from

the naive participants, while the bar to the right shows

the result from the informed participants.

The results show that 62.5% of the naive partic-

ipants performing a task reported a correct result in

the HQ/TWFQ condition. When the participants were

informed about assessing rendering quality, 56.25%

reported a correct result in the same condition while

performing a task.

The percentages for the naive and informed partic-

ipants free-viewing the same condition were 56.25%

and 93.75% respectively. While the correct classiﬁca-

tion frequency increased for the informed participants

free-viewing the scene, the level when performing a

task while being informed was not altered greatly.

4.5 Statistical Analysis and Discussion

The results were analysed statistically using the Chi-

square test to determine any signiﬁcance. This test al-

lows us to determine if what is observed in a distribu-

tion of reported frequencies (correct/incorrect) would

be what is expected to occur by chance. The reported

frequencies were compared to an expected 50/50 data

to ascertain whether the participants perceived the dif-

ference in quality.

The statistical analysis of the results conﬁrmed

that for naive participants performing a task in the

HQ/TWFQ condition, the difference in proportions

was not signiﬁcant χ

(1, N = 16) = 1, p = 0.32. The

results were also not signiﬁcant for the informed

participants that performed a task and the naive

participants free-viewing the two animations, both

(1, N = 16) = 0.25, p = 0.62. This indicates that

the two different animations were perceived as a sim-

ilar quality in these three conditions.

The results for the participants free-viewing the

two animations while being informed differed from

the other three conditions. In this case, almost all par-

ticipants managed to perceive the higher quality ani-

mation, χ

(1, N = 16) = 6.25, p = 0.01. This could

potentially be caused by the fact that they were not

just free-viewing the scene anymore.

GRAPP 2008 - International Conference on Computer Graphics Theory and Applications

212

HQ/TWFQ - naive HQ/TWFQ - informed

Animation Conditions

Num. of Preferences

Performing Task

Correct

Incorrect

HQ/TWFQ - naive HQ/TWFQ - informed

Animation Conditions

Num. of Preferences

Free-viewing

Correct

Incorrect

Figure 4: Results from the no foveal region experiment for the two conditions: (left) performing a task (counting ﬁre safety

items) vs. (right) watching the animations.

Asking people to watch the animations, while be-

ing informed about assessing rendering quality, might

have constituted an implicit task in itself and poten-

tially made them focus on parts of the scene where

quality differences were most likely to occur, for ex-

ample edges.

Mack and Rock (Mack and Rock, 1998) observed

that conscious perception requires attention, which

agrees with the obtained results. The participants in

the new experiment failed to notice the low render-

ing quality within the foveal region of the animations.

This indicates that rendering quality can be reduced

within the foveal region and that IB can in fact be ex-

ploited in selective rendering.

5 CONCLUSIONS

This paper presented a psychophysical forced-choice

preference experiment assessing if participants, per-

forming a task or free-viewing animations, would

fail to notice rendering quality degradation within the

foveal region. Previous work in task-related render-

ing assumed that IB would be restricted to the area

outside the foveal region. However, IB studies have

shown that observers conducting a task can fail to see

an object although it coincides with a ﬁxation. The

experiment presented in this paper showed for the ﬁrst

time that IB can be exploited in selective rendering,

by using a sharp cut-off around the task objects, while

maintaining a perceptually high quality result.

The only observers to show a statistically signiﬁ-

cant ability to detect the quality difference were those

who were free-viewing the HQ/TWFQ pairs after be-

ing informed that they would be asked to judge the

quality of the animations. This is an interesting result

which agrees with the ﬁndings in (Mack and Rock,

1998) which propose that there is no conscious per-

ception without attention.

The experiment indicated that reductions in qual-

ity can be achieved through analysis of high-level vi-

sual processing, and is not bound to low-level bio-

logical processes, such as knowledge of the foveal re-

gion. This is an improvement on previous work which

added a foveal region around the task objects. By be-

ing able to reduce the quality further than previously

proposed ROI rendering techniques, additional com-

putation savings can be made.

Overall, the experiment shows that there is an op-

portunity in ROI rendering, but that more research

must be done to better understand how rendering

quality affectshuman perception. It could also be pos-

sible that using participants which are familiar with

artifacts can lead to an overly strict criteria. Hence, as

future work it would be interesting to study if there are

signiﬁcant differences between participants being fa-

miliar with the concept of rendering quality and those

who are not. As future work it would also be inter-

esting to study the eye movements of the participants

while they make the quality judgements.

ACKNOWLEDGEMENTS

The author would like to thank Kurt Debattista for

implementation of ROI rendering in Radiance, Pe-

ter Longhurst for the modiﬁed Snapshot program

and Sumaya Ahmed for collaboration on the task

map plug-in. Thanks also to Andrew Moss, Alan

Chalmers, everyone that participated in the experi-

A PSYCHOPHYSICAL STUDY OF FOVEAL GRADIENT BASED SELECTIVE RENDERING

213

ments, and Patrick Ledda for the original corridor

model. This research was partly done under the spon-

sorship of the Rendering on Demand (RoD) project

within the 3C Research programme and by the Euro-

pean Union within the CROSSMOD project (EU IST-

014891-2).

REFERENCES

Baudisch, P., DeCarlo, D., Duchowski, A. T., and Geisler,

W. S. (2003). Focusing on the essential: considering

attention in display design. Commun. ACM, 46(3):60–

66.

Brown, R., Cooper, L., and Pham, B. (2003). Visual

attention-based polygon level of detail management.

In GRAPHITE ’03: Proceedings of the 1st interna-

tional conference on Computer graphics and interac-

tive techniques in Australasia and South East Asia,

pages 55–62, New York, NY, USA. ACM Press.

Cater, K., Chalmers, A., and Ledda, P. (2002). Selec-

tive quality rendering by exploiting human inatten-

tional blindness: looking but not seeing. In VRST ’02:

Proceedings of the ACM symposium on Virtual real-

ity software and technology, pages 17–24, New York,

NY, USA. ACM Press.

Cater, K., Chalmers, A., and Ward, G. (2003). Detail to at-

tention: exploiting visual tasks for selective rendering.

In EGRW ’03: Proceedings of the 14th Eurograph-

ics workshop on Rendering, pages 270–280, Aire-la-

Ville, Switzerland, Switzerland. Eurographics Associ-

ation.

Debattista, K. (2006). Selective Rendering for High-

Fidelity Graphics. PhD thesis, University of Bristol.

Haber, J., Myszkowski, K., Yamauchi, H., and Seidel, H.-

P. (2001). Perceptually guided corrective splatting.

Computer Graphics Forum, 20(3):142–152.

Howlett, S., Hamill, J., and O’Sullivan, C. (2005). Predict-

ing and Evaluating Saliency for Simpliﬁed Polygonal

Models. ACM Trans. Appl. Percept., 2(3):286–308.

Itti, L., Koch, C., and Niebur, E. (1998). A Model

of Saliency-Based Visual Attention for Rapid Scene

Analysis. IEEE Trans. Pattern Anal. Mach. Intell.,

20(11):1254–1259.

James, W. (1957). The Principles of Psychology. Dover

Publications Inc.

Lee, C. H., Varshney, A., and Jacobs, D. W. (2005). Mesh

saliency. In SIGGRAPH ’05: ACM SIGGRAPH 2005

Papers, pages 659–666, New York, NY, USA. ACM

Press.

Longhurst, P. (2005). Rapid Saliency Identiﬁcation for Se-

lectively Rendering High Fidelity Graphics. PhD the-

sis, University of Bristol.

Loschky, L. C., Mcconkie, G. W., Reingold, E. M., and

Stampe, D. M. (2003). Gaze-Contingent Multiresolu-

tional Displays: An Integrative Review. Human Fac-

tors, 45(2):307–328.

Mack, A. and Rock, I. (1998). Inattentional Blindness. MIT

Press.

McConkie, G. W. and Loschky, L. C. (1997). Human

Performance with a Gaze-Linked Multi-Resolutional

Display. In Advanced Displays and Interactive Dis-

plays First Annual Symposium, pages 25–34.

McNamara, A. (2001). Visual Perception in Realistic Image

Synthesis. Computer Graphics Forum, 20(4):211–

224.

Myszkowski, K., Przemyslaw, R., and Tawara, T.

(1999). Perceptually-informed Accelerated Render-

ing of High Quality Walktrough Sequences. In Euro-

graphics Workshop on Rendering, pages 13–26. The

Eurographics Association.

O’Sullivan, C. (2005). Collisions and Attention. ACM

Trans. Appl. Percept., 2(3):309–321.

O’Sullivan, C., Howlett, S., Morvan, Y., McDonnell, R.,

and O’Conor, K. (2004). Perceptually Adaptive

Graphics. In Eurographics 2004, STAR, pages 141–

164. The Eurographics Association.

Snowden, R., Thompson, P., and Troscianko, T. (2006). Ba-

sic Vision: an introduction to visual perception. Ox-

ford University Press.

Sundstedt, V. (2007). Rendering and Validation of Graph-

ics using Region-of-Interest. PhD thesis, University of

Bristol.

Sundstedt, V., Debattista, K., Longhurst, P., Chalmers, A.,

and Troscianko, T. (2005). Visual attention for efﬁ-

cient high-ﬁdelity graphics. In SCCG ’05: Proceed-

ings of the 21st spring conference on Computer graph-

ics, pages 169–175, New York, NY, USA. ACM Press.

Ward, G. J. (1994). The RADIANCE lighting simula-

tion and rendering system. In SIGGRAPH ’94: Pro-

ceedings of the 21st annual conference on Computer

graphics and interactive techniques, pages 459–472,

New York, NY, USA. ACM Press.

Woolley, C., Luebke, D., Watson, B., and Dayal, A. (2003).

Interruptible rendering. In SI3D ’03: Proceedings

of the 2003 symposium on Interactive 3D graphics,

pages 143–151, New York, NY, USA. ACM Press.

Yang, X. and Chalmers, A. (2005). Perceptually Driven

Level of Detail for Efﬁcient Ray Tracing of Com-

plex Scenes. In Proceedings of Theory and Practice

of Computer Graphics 2005, pages 91–96. The Euro-

graphics Association.

Yarbus, A. L. (1967). Eye movements during perception of

complex objects. In Eye Movements and Vision, pages

171–196, New York. Plenum Press.

Yee, H., Pattanaik, S., and Greenberg, D. P. (2001). Spa-

tiotemporal sensitivity and visual attention for efﬁ-

cient rendering of dynamic environments. ACMTrans.

Graph., 20(1):39–65.

GRAPP 2008 - International Conference on Computer Graphics Theory and Applications

214