DtMF: Debí… ¿tirar más fotos? tener Más Fijaciones — Modeling Human Variability and Same-Image Target-Presence Comparisons in Visual Search

Computational models of visual search have become increasingly successful at predicting observers’ fixation patterns during free viewing. A recent approach combining deep neural networks for bottom-up information processing with Bayesian top-down integration achieved the best performance across many datasets in the ViSioNS benchmark. However, two important limitations remain. First, comparisons between target-present and target-absent search are often based on different sets of images, making it difficult to disentangle effects of target presence from differences in scene content. To address this, we built a new dataset and framework in which target-present and target-absent trials can be directly compared under matched visual conditions. Second, existing metrics focus on average performance, although observers’ scanpaths in the same scene are not identical. We therefore investigated whether variability across observers could be reproduced by introducing structured stochasticity in the initial prior (saliency map) and in the expected-information map at each fixation. Across different noise parameters, introducing internal noise not only captured observer variability but also improved model performance. We conclude that stochasticity is inherent to human information processing and propose that future models should be evaluated both by their average performance and by their ability to reproduce human variability.