Integration of local and global cues to reconstruct surface structure

Abstract

The computation of border-ownership (BOWN) and the reconstruction of surface structure - i.e., figure/ground assignment and the interpolation of missing contours - are essential puzzles of visual computation, in part because they epitomize the integration of local and global cues to generate a coherent percept. Here we attempt to integrate two previous computational models and bring them to bear on this problem. In the DISC model of Kogo et al. (2010), BOWN was computed by global iterative interactions of image elements, BOWN signals were considered as a differentiated form of surface representation, and 2-D integration of BOWN signals (re-) constructed surfaces. Furthermore, the perception of illusory contours and surfaces was modeled by assuming that there are potential BOWN signals at every location in the entire space (free-space BOWNs). In a complementary fashion, Froyen et al. (2010) showed how BOWN can be estimated using Bayesian belief propagation, integrating both local cues (e.g., T-junctions and sign of curvature) and global ones (e.g., skeletal structure) in principled ways. This model included as a nonlocal factor skeletal (medial-axis) structure, under the hypothesis that the medial structure that explains the border best draws its ownership. Here we combine these approaches to yield estimates of surface structure throughout the image, including both the interiors of surfaces as well as all points along the boundaries. We integrate the idea of free-space BOWN to include the computation of illusory contours into the Bayesian framework. Within this dynamic generative model, free-space BOWN signals are estimated by recurrent feedback from higher-level medial structure. Two processes alternate iteratively to estimate local free-space BOWN: (1) skeletal structure is estimated from the BOWN signals and (2) skeletal structure generates new free-space BOWN signals. This process eventually converges onto estimates that are in line with human perception.