Congruence between model and human attention reveals unique signatures of critical visual events
Robert J. Peters and Laurent Itti ~ NIPS*2007 Poster/Spotlight ID W30
(1) We record the eye positions of people playing video games; we isolate critical events, like missile firings in a flight combat game (2) Computational models produce maps of likely eye positions, based on bottom-up salience and top-down relevance (3) We quantify ("NSS") how well each model map agrees with the eye position, and find unique NSS signatures during critical events (4) These NSS signatures can be used to non-invasively predict what players are seeing and doing li 1.5s before missive-ftring e en 1.5s after

1

NSS

Salience vs. Eye pos.
-5 -1.5 0 1.5

0

-10

Relevance vs. Eye pos.
5

time (s) relative to missile-firing event

10