# Ocular drift shakes the stationary view on pattern vision

## TL;DR

This study shows that eye movements, like ocular drift, affect how we perceive edges in visual scenes, challenging traditional models of vision.

## Contribution

The study reveals that current models of spatial vision implicitly compensate for the absence of eye movements, suggesting a need to revise traditional assumptions.

## Key findings

- Incorporating ocular drift into a mechanistic model of spatial vision surprisingly led to worse performance compared to the original model.
- A simpler model with a single spatial frequency channel benefits from drift but performs poorly without it.
- Standard models of spatial vision may favor a stationary view of input, potentially leading to self-confirming theories.

## Abstract

The mechanisms by which the visual system extracts key features (i.e., edges) from the visual input remain not fully understood. As reflected in the term spatial vision, pattern vision is traditionally assumed to operate on stationary visual inputs. However, our eyes are never truly still. Involuntary eye movements, specifically ocular drift, continuously alter the visual input during fixations and redistribute its power, emphasizing high spatial frequency contents. In this study, we examine the role of ocular drift on edge sensitivity in noise. We show that drift-induced shifts in stimulus power lead to better predictions of the empirical data, consistent with the human contrast sensitivity function. We then incorporate drift into a mechanistic model of spatial vision to test whether this further improves model predictions. Surprisingly, the original spatial model outperforms the drift-enhanced version. It does so in an interesting way: It artificially compensates for the absence of drift by redistributing the activity across its spatial frequency channels in later processing stages, effectively mimicking the effect of a dynamic input without explicitly modeling it. By contrast, a simpler model with a single spatial frequency channel benefits from drift but performs poorly when drift is removed. These findings suggest that standard model architectures inherently favor a stationary view of visual processing, which could result in self-confirming theories. Incorporating the dynamic nature of the visual input may offer a more accurate model of how the brain processes key features of natural scenes. However, doing so requires a critical reassessment of long-standing frameworks in visual neuroscience.