Visual segmentation is a core function of biological vision, key to adaptive behavior in complex environments. Foundational work identified Gestalt principles of segmentation, e.g. grouping by similarity, proximity and good continuation, and revealed that visual cortical neurons are sensitive to those cues. Early models inspired by the feedforward cortical architecture described texture-based human segmentation as the process of comparing the summary statistics of low-level visual features across space. Indeed the summary statistics representation is the most prominent model of naturalistic texture perception, yet it has been challenged precisely because it does not fully capture the influence of segmentation.
Here we consider the alternative view that, due to image ambiguity and sensory noise, perceptual segmentation requires probabilistic inference. This view is consistent with reports that humans combine multiple segmentation cues near-optimally in artificial displays, and that Gestalt laws reflect optimization to natural image statistics. The probabilistic approach is also widespread in computer vision algorithms for unsupervised segmentation, but has not been used to model perceptual segmentation. We present new experiments that for the first time measure perceptual segmentation maps and their variability, allowing us to test the probabilistic inference hypothesis, and compare it quantitatively to summary statistics models.
We use composite textures, with segments characterized by different statistical relations between features. Optimal probabilistic inference assigns pixels to segments by evaluating which of those relations better explains the observed features (generative model), as opposed to comparing summary statistics at different locations (feature discrimination). We find the generative model best captures our data, and perceptual variability reflects image uncertainty beyond sensory noise. We also demonstrate the approach on natural images, which will allow testing more sophisticated segmentation algorithms. Our results provide a normative explanation of human perceptual segmentation as probabilistic inference, and demonstrate a novel framework to study perceptual segmentation of natural images.