Abstract
Studies of visual search—looking for targets among distractors—typically focus on quantifying the impact of general factors (e.g., number of distractors) on search performance. However, search efficiency, particularly in complex environments, is undoubtedly a function of the particular similarity relationships between the specific target(s) and distractors. Further, the visual system represents many different dimensions (color, location, category) that can be flexibly weighted according to current goals, implying that different similarity relationships may be important in distinct contexts. In the current study we examined the impact of similarity in complex visual search by using “big data” from the mobile app Airport Scanner, where the player serves as an airport security officer searching bags for a diverse set of prohibited items among a large heterogeneous set of potential distractors. This large variability in possible targets and distractors, combined with the volume of data (~3.6 billion trials, ~14.8 million users), provide a means to explore the impact of target-distractor similarity on search. The game also includes levels that players advance through in sequence, enabling an investigation of the effect of experience. The data were used to calculate the impact of every distractor on every target at each level, and the resulting behavioral matrices were then compared to a number of different similarity metrics derived from image statistics (e.g., color, pixelwise) and biologically-inspired models of vision (e.g., HMAX). The analyses revealed that experience shaped the impact of distractors, with lower-level metrics dominant early and higher-level features becoming increasingly important as target and distractor familiarity increased. The detailed understanding of search revealed by these analyses provides key insights for generating a detailed model of real-world search difficulty.
Acknowledgement: Army Research Office