Standard models of spatial vision mispredict edge sensitivity at low spatial frequencies.

Area of Science:

• Visual neuroscience
• Perceptual psychology
• Computational vision

Background:

• The contrast sensitivity function (CSF) is a key model of early visual processing, describing sensitivity across spatial frequencies (SF).
• Standard spatial vision models, based on CSF, use SF-selective channels but are less tested with naturalistic stimuli like high-contrast edges.
• Understanding sensitivity to edges is crucial for explaining object boundary perception in natural environments.

Purpose of the Study:

• To evaluate the efficacy of standard spatial vision models in predicting sensitivity to sharp edges.
• To investigate edge sensitivity under broadband and narrowband noise conditions.
• To identify deviations from model predictions that may reveal limitations or new aspects of visual processing.

Main Methods:

• Edge localization tasks (2-alternative forced choice) were used to measure sensitivity to Cornsweet luminance profiles.
• Stimuli featured peak spatial frequencies of 0.5, 3, and 9 cycles per degree (cpd).
• Single- and multi-scale computational models, derived from standard spatial vision components, were implemented to predict human performance.

Main Results:

• Both implemented models largely accounted for edge sensitivity data across different spatial frequencies and noise types.
• Systematic deviations were observed, particularly for the lowest spatial frequency (0.5 cpd) edge and in the presence of pink noise.
• Model predictions diverged from empirical data, indicating areas where current spatial vision models may be insufficient.

Conclusions:

• Current spatial vision models offer a good approximation but do not fully capture edge sensitivity, especially at low spatial frequencies.
• Deviations suggest a possible shift from contrast-based to luminance-based detection at low SFs or the need for additional components in spatial vision models.
• Further research is needed to refine models for predicting perception of behaviorally relevant stimuli like object boundaries.