cyclical vs. stochastic visual search models andrew elias 11 aug 2011 nrl code 7440 branch meeting...
TRANSCRIPT
Cyclical vs. Stochastic Visual Search Models
Andrew Elias11 Aug 2011NRL Code 7440 branch meeting
How a Target’s Effect on the Eye Varies with Distance
&
Eye TrackingFixations: eye stops (avg. 400 ms)
Saccades: jerky eye movements, no data intake
The Experiment
30 FA18 pilots
72 images24 distinct mapsWith/without targetssorted:
3 clutter levelsHigh/low local clutter near target
Target:
Cyclical vs. Stochastic Visual Search Models
Cyclical: avoid re-examining things
Stochastic: don’t worry about re-examining things
Disclaimer: Examination Radius (parameter) chosen to make Stochastic model fit the best
…now choose parameters to make Cyclical model fit the best
Conclusion
Stochastic model is correct.(at least for the majority of our trials)
Further Research
• Comparing clutter levels or pilot experience levels: do models/parameters change?
• Using model parameters as a new measure of global clutter (!!!) or pilot skill
• Analyze subjects/maps individually to see if some use/allow Cyclical model.
How a Target’s Effect on the Eye Varies With
Distance
Purpose: fixation “size”
small? large? fuzzy?
How long are saccades?
At what distance does saccade length change?
120 px
55 px
Map Edge
Successful Saccade (n.)
a saccade which ultimately brings
the subject closer to the target
Seeing or remembering
P(S
ucc
ess
)
P(S
ucc
ess
| N
ot
Seein
g)
Distance to Target (px) Distance to Target (px)
Distance to Target (px)
P(S
eein
g)
“Bayesian” derivation of P(Seeing)
Successful Run-of-Saccades
L: the number of saccades into the future that we observe
Conclusions
• 55 px is the threshold for glance-away saccades
• Attention-grabbing radius may be fuzzy (at least for a large group), but probably spans a subset of the the 55-450 px zone. Maybe ~120 px
• These values may be target-dependent and map-dependent.– ‘450 px’ value is especially map-dependent.
• Reaction “time” is less than two saccades.
a visualization:
44 (±3) px/deg
Further Research
• Remove all after-target fixations and see if 2nd bump disappears, and if 1st bump gets clearer
• Blend distributions to model the control-group function
Acknowledgments
• Maura Lohrenz, 7441 section head• Melissa Beck (LSU)• Todd, Stephanie, Mike, Jeremy• fellow students!• experiment participants
somebody ask a question!
The end
Some initial readings…