Learning from nature: image segmentation based on local symmetry detection

Symmetry has become synonymous with abstract shape representation in human vision. However, recordings from neurons in low-level visual brain areas in monkeys revealed neural sensitivity to medialness. The same neurons also detect edges and oriented lines. We propose a symmetry computation scheme in noisy images, and re-interpret the role of symmetry in vision to understand why it is important at early stages in the hierarchy of the primate visual system. The new symmetries correspond closely to that observed in human subjects. Noise sensitivity can be controlled, overcoming severe restrictions on traditional models. Boundary reconstruction leads to meaningful segmented image regions as illustrated with synthetic images, and are useful in tasks like detecting boundaries in textures. Model performance and problems encountered with natural images are demonstrated. Extension to 3D does not introduce additional complications