Looking for Image Statistics: Active Vision With Avatars in a Naturalistic Virtual Environment

Area of Science:

• Computational Neuroscience
• Visual System Modeling
• Image Statistics

Background:

• The efficient coding hypothesis suggests sensory systems match natural input statistics.
• Previous studies on natural image statistics often used camera-based datasets, neglecting visual system properties.
• Active vision, involving body and eye movements, may alter image statistics.

Purpose of the Study:

• To quantitatively investigate how the active use of the visual system influences image statistics across the visual field.
• To determine if simulating visual behaviors in a virtual environment can reveal these influences.
• To compare image statistics generated by active vision with those from traditional photographic datasets.

Main Methods:

• Simulated visual behaviors of human and cat avatars in a virtual forest environment.
• Generated images with a 120° field of view, projected onto idealized retinas.
• Analyzed image statistics, including power spectra and orientation biases, under different active gaze behaviors.

Main Results:

• Central visual field statistics matched photographic images (power spectra, cardinal orientation bias).
• At larger eccentricities, a radial bias emerged, influenced by active behavior and eye optics.
• Significant differences were found between upper and lower visual fields, varying with environmental sampling.

Conclusions:

• Relating natural image statistics to neural and behavioral data requires considering the environment, visual system physics, and active use.
• Active vision and eye geometry introduce biases in image statistics, particularly at peripheral visual field locations.
• Simulations provide a valuable method for studying the impact of active visual behaviors on sensory input statistics.