Abstract
Human detection performance in natural backgrounds, as a function of the luminance, contrast, and similarity of the background to the target, is predicted directly from the statistics of these properties in natural images (Sebastian et al., PNAS, 2017). Another potentially important factor for detection in natural backgrounds is that these properties of the background may vary under the target, and hence the target may not be uniformly masked by the background (i.e., the masking may only partially occlude the target). We’ve previously demonstrated that a reliability-weighted template-matching model, which estimates local reliability from the local contrast, predicts human thresholds for detecting targets in spatially modulated white noise (Sebastian & Geisler, JOV abstract, 2017). The current study assessed trial-by-trial correlations in a similar target detection task in both white noise and natural backgrounds. In one condition, subjects detected targets of different amplitudes in white noise backgrounds, where the noise contrast modulated under the target. In a second condition, the backgrounds were natural scenes that were chosen to have local variation in contrast. For both conditions, decision-variable correlations (Sebastian & Geisler, JOV, 2018) were computed to assess the trial-by-trial correlations between the human observers and two models: one with and one without local reliability weighting. In both conditions, the reliability-weighted template model better predicted human responses than the simple template model, for both target-present and target-absent trials. These results suggest that the visual system uses local reliability weighting when detecting targets in complex backgrounds.