An Experimental Approach to Predicting Saliency for Simplified Polygonal Models

1
An Experimental Approach to Predicting Saliency
for Simplified Polygonal Models

• Authors Sarah Howlett, John Hamill and Carol
  OSullivan
• Affiliation Image Synthesis Group
• My e-mail address Sarah.Howlett_at_cs.tcd.ie

2
The old SMI EyeLink Eye-Tracking Device
The new EyeLink 2 Eye-Tracking Device
The eye-tracker was used to ascertain prominent
features of models
3
Simplified to various levels of detail (LOD)-
Full LOD 3700 POLYGONS
Original
Modified
50 LOD
20 LOD
5 LOD
2 LOD
(i.e. 1850 Polygons)
4
(No Transcript)
5
Introduction

• The ideal is to have a realistic dynamic scene
  within real time constraints.
• The challenge is maintaining the appearance of
  the models under simplification.
• Perceptually adaptive graphic taking advantage
  of the human visual system.

6
Background

• Geometry - Garland and Heckbert 97. Rushmeier
  01.
• Perceptual models - Luebke and Hallen 01.
• Input taken from the user - Kho and Garland 03.
  Pojar and Schmalstieg 03. Proceeded by Cignoni
  et al. 98. Li and Watson 01.
• Gaze contingent systems Duchowski 02
• Peripherally degraded displays Reddy98.
  Watson97.
• Saliency Itti et al. 98. Yee et al. 01

7
Finding Salient Features

• We gathered information on where a participant
  was fixating while viewing a models.
• Three metrics were used.
• This experiment was carried out on two sets of
  models.

8
The total length of fixations on each face
Saliency color map ranging from red through to
yellow, green, cyan, blue, magenta and finally
white.
more
less
9
The duration of the first fixation on each face
The total number of fixations on each face
more
less
10
(No Transcript)
11
Evaluation

• We incorporated fixation data to produce a
  modified Quadric error metric
• This was applied to the QSlim software developed
  by Garland and Heckbert.

5 original
5 modified
12
Finding the Naming Times

• Naming time experiments were carried out on the
  set of familiar objects.
• The name of objects within the category would be
  known to the participant e.g. car, elephant or
  chair.
• We recorded the naming time and the number of
  errors for models simplified using original and
  modified QSlim.

2 original
2 modified
13
Results

• Results only affected by simplification level at
  low LODs.

• At full LOD it took longer to name natural
  objects.
• At 2 LOD natural objects simplified using
  modified QSlim were named faster.

14
(No Transcript)
15
Acquiring the Picture-Picture Matching Times

• We examined further categorical effects on the
  second set.
• For unfamiliar object subjects were not expected
  to know or even remember the name of individual
  objects within a category.
• The matching time and the number of correctly
  matched objects were recorded.
• The 4 categories of model animals, cars, fish
  and gears were simplified to various LODs.

16
Results

• At low LODs results were affected by
  simplification level.
• Animals were the fastest matched. Cars were
  matched slowest.

• For the animal model at 5 and 2 results were
  significantly better when modified QSlim was
  used.
• Also the modified fish at 30 were named more
  accurately.

17
(No Transcript)
18
Forced-Choice Preference

• To investigate further forced-choice preference
  tests were carried out on both sets of models.
• A web-based interface was used for this.
• All models under the two simplification types
  were compared at the same LOD.

19
Results

• In the first online experiment there was a
  preference for the modified natural objects and
  for the original man-made artifacts.
• For the unfamiliar objects there was a preference
  for the modified fish objects.

20
(No Transcript)
21
Conclusions

• There were prominent features in the case of some
  models.
• We found promising results for the natural
  objects at low LODs.
• Saliency based simplification can enhance the
  visual fidelity of natural objects at low LODs.

22
Future research

• To examine better approaches to saliency
  determination for synthetic objects.
• Are the prominent features more defined by a
  specific task.
• Need ideas for experiments to examine and
  evaluate this further.

23
(No Transcript)
24
References

• Garland and Heckbert. Surface simplification
  using quadric error metrics. 1997
• Kho and Garland. User-guided simplification.
  April 2003
• Watson, Friedman and McGaffey. Measuring and
  predicting visual fidelity. 2001
• Lawson, Bulthoff and Dumbell. Interactions
  between view changes and shape changes in
  picture-picture matching. May 2002
• Luebke, Hallen, Newfield and Watson. Perceptually
  Driven Simplification Using Gaze-Directed
  Rendering. 2000
• Pojar and Schmalstieg. User-Controlled Creation
  of Multiresolution Meshes. April 2003
• Hayhoe. Vision using routines A functional
  account of vision. 2002