Hierarchical, Localized Saliency Computations Solve the Visual Feature Binding Problem

Computational vision has a long history of proposing methods for decomposing a visual signal into components. What has been far more elusive is how to recombine those components into a whole, a problem known as the binding problem. Although several proposals have appeared, the approaches and their demonstrations seem weak at best. This paper proposes a novel solution for a significant portion of the binding problem, namely, the re-combination of visual features into larger patterns and their localization in the image. The solution requires the abandonment of the nearly ubiquitous single saliency map and the adoption of a hierarchical, localized computation of saliency that is dependent on local neural selectivity constraints. This strategy has been demonstrated within a fully implemented model that attends to simple motion patterns in image sequences. 1