Do People Take Stimulus Correlations into Account in Visual Search?

Abstract

In laboratory visual search experiments, distractors are often statistically independent of each other. However, stimuli in more naturalistic settings are often correlated and rarely independent. Here, we examine whether human observers take stimulus correlations into account in orientation target detection. We find that they do, although probably not optimally. In particular, it seems that low distractor correlations are overestimated. Our results might contribute to bridging the gap between artificial and natural visual search tasks.

Fig 1. Experimental procedure and sample displays.

(a) Time course of a trial, (b) Sample displays for each of the correlation coefficients used. In a given experimental session, only one value of ρ was used.

Fig 2. Psychometric curves 1.

(a) Proportion correct responses and (b) hit and false alarm rates as a function of distractor correlation. Throughout the paper, error bars indicate one standard error of the mean (s.e.m).

Fig 3. Psychometric curves 2.

(a) Proportion “target present” responses and (b) number of trials as a function of standard deviation of the distractor set coefficient, averaged across subjects. Bin size was 2.5°, except that all trials with sample standard deviation greater than 17.5° are collected in the last bin. The plots in (b) are entirely determined by the stimuli, not by the subject responses; they serve to emphasize that the points in the plots in (a) were computed on widely differing numbers of trials.

Fig 4. Psychometric curves 3.

(a) Proportion “target present” responses and (b) number of trials as a function of minimum target-distractor orientation difference for target present (left) and absent (right) trials. Bin size was 2°, except that all trials with minimum target-distractor orientation difference greater than 10° are collected in the last bin. The plots in (b) are entirely determined by the stimuli, not by the subject responses; they help to reconcile the plots in (a) with Fig 2b.

Fig 5. AIC model comparison for equal versus variable precision.

Shown are AIC differences of EP models relative to VP models for each subject (left) and averaged over subjects (right). Higher AIC mean worse fits. BIC results are consistent (S1a Fig).

Fig 6. AIC model comparison of VP models for observer’s assumption about ρ and parameter estimates of VP4 model ρ_assumed.

(a) Shown are AIC differences of VP models relative to VP4 (most general) model for each subject (left) and averaged across subjects (right). (b) ML estimates of ρ_assumed from the VP4 model for each subject (colors) and averaged (black). BIC results are consistent (see S1b Fig).

Fig 7. Fits of the VP4 model to the summary statistics.

(a) Proportion correct (top), hit, and false-alarm rates (bottom) as a function of distractor correlation. Proportion “target present” responses as a function of (b) standard deviation of the distractor set, and (c) minimum target-distractor orientation difference, averaged across subjects, separately for target present (black) and target absent (red) trials. Numbers indicate root-mean square error (blue) and R² statistics (green) between model and data.

Fig 8. Fits of the VP1 model to the summary statistics.

For caption, see Fig 7.

Fig 9. Fits of the VP5 model to the summary statistics.

For caption, see Fig 7.

Fig 10. AIC model comparison of VP models relative to VP5 model and parameter estimates of VP5 model.

(a) Shown are AIC differences of VP models relative to VP5 model for each subject (left) and averaged across subjects (right). BIC results are consistent (S1c Fig). (b) ML estimates of $\overline{J}$ and ρ_assumed from the VP5 model for each subject (colors) and averaged with standard error mean across subjects (black).