Natural scene segmentation dynamics reveal iterative Bayesian inference

Abstract

The visual system operates by segmenting visual inputs into distinct perceptual objects. Segmentation is dynamic, as revealed by the tempo of perceptual choices and neural activity in the visual cortex. Dynamics for natural stimuli however, are poorly understood because natural scene segmentation is ambiguous and subjective. We measured subjective human segmentation maps for natural images using an innovative paradigm, and uncovered richer spatiotemporal dynamics than predicted by current theories of segmentation. To explain these dynamics, we introduced Iterative Bayesian Inference algorithms for segmentation that iteratively integrate visual inputs with the prior expectation for local smoothness. When visual inputs were consistent with such a spatial prior, iterative inference was faster. Such a predicted relationship between spatial prior and inferential dynamics was evident in our data and correctly reflected each individual participant's spatial biases. We conclude that iterative Bayesian inference sets the tempo for a fundamental function of natural vision.