Visual search on natural images: Incorporating noise as a model of human variability

Abstract

Visual search is crucial in daily human interaction with the environment, involving sequential eye movements. Computational models attempt to replicate this behavior by mimicking cognitive mechanisms used during natural image perception. Despite advances in neuroscience, visual search remains complex and not fully characterized. One key aspect of human cognition often overlooked is the role of noise, particularly in exploration and decision-making. Different sources of noise cause variability in response times, provoke mistakes, and may also explain individual differences between participants. We propose a modified version of the Entropy Limit Minimization (ELM) model to better capture human variability in visual search. Although deterministic, our model introduces randomness through seeds derived from both the input image and the participant. We explore several noise injection strategies to the model, including adding noise to the image prior, the information gain map, and the overshooting of saccades via random offsets. To evaluate variability, we use metrics from the VISIONS benchmark, repeating model runs with different seeds. Notably, adding noise in the prior of the model yielded the closest approximation to human variability. However, the results show that adding different types of noise helps the model  approximate human variability in most cases.