Keynotes | K&Ts | GACs | Talks | Posters | Search

Poster A in Poster Session A: Tuesday, August 4, 9:30 – 11:15 am, Kimmel Center, Shorin & Rosenthal Rooms

Manipulating human gaze reveals a cost structure of natural visual search

Hyunwoo Gu¹, Justin L Gardner¹; ¹Stanford University

Presenter: Hyunwoo Gu

A hallmark of human intelligence is the balance between goal-driven control and more general processes. Visual search is a canonical task in which both systems guide behavior: goal drive prioritizes target-relevant features, while salience drive directs attention toward conspicuous regions. In classical paradigms, these systems can be placed in conflict, reducing efficiency. However, in natural scenes, salience and target relevance may exhibit complex statistical relationships, raising the question of how they jointly guide attention. Here, we manipulate human overt attention in naturalistic visual search and quantify the resulting cost structure. Using deep image synthesis, we generate stimuli with controlled salience and show that this manipulation biases gaze during search. We then characterize search cost using image-computable measures and track how attentional priorities evolve over fixation sequences. We find that (i) gaze can be causally steered by salience, (ii) both effective set size and salience-target conflict increase the search cost, and (iii) a transition from salience- to target-driven guidance over time may underlie this cost.

Topic Area: Computational Models of Vision & Visual Cortex