Natural scene statistics predict how humans pool information across space in surface tilt estimation

Area of Science:

• Vision science
• Computational neuroscience
• Perception

Background:

• The human visual system reconstructs 3D scene structure from 2D retinal input.
• Statistical regularities in natural scenes inform visual perception and improve depth estimation accuracy.

Purpose of the Study:

• To investigate the spatial relationships of 3D surface tilts in natural scenes.
• To determine how humans pool spatial information for surface tilt estimation.
• To develop and test a hierarchical model of surface tilt perception grounded in scene statistics.

Main Methods:

• Analysis of spatial statistics of surface tilts in natural scenes and images.
• Development of a hierarchical computational model for surface tilt estimation.
• Comparison of model performance against human behavioral data and local estimation methods.

Main Results:

• 3D surface tilts exhibit predictable spatial relationships in natural environments.
• A hierarchical model pooling local tilt estimates within an adaptive spatial neighborhood accurately predicts groundtruth tilt.
• The model demonstrates superior performance compared to local estimates and better accounts for human perception.

Conclusions:

• The human visual system leverages spatial statistics of natural scenes to pool surface tilt information adaptively.
• This spatial pooling mechanism enhances the accuracy of 3D surface tilt perception.
• The findings support a hierarchical Bayesian approach to understanding visual surface perception.