Towards an optimal decision strategy of visual search

TL;DR

This paper investigates the optimal decision strategies in visual search tasks, proposing a POMDP-based model and a scalable conjecture, revealing that existing mechanisms are sub-optimal and that scaled diffusions are better suited.

Contribution

It introduces a POMDP framework for modeling optimal visual search strategies and proposes a scalable conjecture applicable to cluttered scenes, improving understanding of decision mechanisms.

Findings

Optimal tradeoff strategy formulated as POMDP solution.

Existing decision mechanisms are sub-optimal compared to the proposed strategy.

Scaled diffusions outperform diffusion-to-bound and maximum-of-output methods.

Abstract

Searching for objects amongst clutter is a key ability of visual systems. Speed and accuracy are often crucial: how can the visual system trade off these competing quantities for optimal performance in different tasks? How does the trade-off depend on target appearance and scene complexity? We show that the optimal tradeoff strategy may be cast as the solution to a partially observable Markov decision process (POMDP) and computed by a dynamic programming procedure. However, this procedure is computationally intensive when the visual scene becomes too cluttered. Therefore, we also conjecture an optimal strategy that scales to large number of clutters. Our conjecture applies to homogeneous visual search and for a special case of heterogenous search where the signal-to-noise ratio differs across location. Using the conjecture we show that two existing decision mechanisms for analyzing human data, namely diffusion-to-bound and maximum-of-output, are sub-optimal; the optimal strategy instead employs two scaled diffusions.