2-D scene analysis using split-level relaxation

We present a new method for applying multiple semantic constraints based on discrete relaxation. A separate graph is maintained for each constraint relation and used in parallel to achieve a consistent labeling. This permits both local and global analysis without recourse to complete graphs. Here local means with respect to a paricular constraint graph, and thus actually includes global spatial relations on the features; e.g., parallel edges on an object will be neighbors in the parallel constraint graph even though they are far apart in Euclidean space. Another major result is a technique for handling occlusion by incorporating the use of spatially local feature sets in the relaxation-type updating method.