Bayesian Modeling of Motion Perception Using Dynamical Stochastic Textures.

Area of Science:

• Computational neuroscience
• Visual perception
• Machine learning

Background:

• Psychophysical biases in vision are often modeled as optimal inference processes.
• Generative models provide a framework for understanding these biases.
• Dynamic texture models offer a method to probe visual motion perception.

Purpose of the Study:

• To formulate a complete dynamic texture model for visual motion perception.
• To investigate the influence of spatial frequency on perceived speed.
• To explain human perceptual biases using a Bayesian inference framework.

Main Methods:

• Developed a dynamic texture model derived axiomatically and constrained by biological plausibility.
• Formulated the model using three equivalent approaches: aggregated warped patterns, stochastic partial differential equations (sPDEs), and autoregressive processes.
• Derived a local motion-energy model for Bayesian inference.
• Conducted psychophysical experiments on human speed perception using stimuli with varying spatial frequencies.

Main Results:

• Human data replicated previous findings: perceived speed is positively biased by spatial frequency increments.
• A Bayesian observer model, incorporating a Gaussian likelihood with spatially dependent width and a slow-speed prior, accurately predicted the observed perceptual bias.
• The bias was explained by a decrease in the observer's likelihood width as spatial frequency increased, aligning with the dynamic texture model's properties.

Conclusions:

• The proposed dynamic texture model provides a robust framework for understanding visual motion perception and associated biases.
• Bayesian inference, utilizing a spatially dependent likelihood and prior, effectively explains human speed perception.
• The findings highlight the interplay between spatial frequency, motion perception, and the underlying neural mechanisms.