Figure 5.

An illustration of border-ownership assignment. (A) Initially, each edge in the image (red arrows) can belong to one of two sides (green arrows): either the gray surface or the white surface. The goal of border ownership computation is to figure out which side is the figural side. In this case, the edges should belong to the gray surface. (B) Using the convexity assumption (objects tend to be convex), we can easily determine border ownership in the local neighborhood. Edges that “agree” (green arrows are “looking” at each other) are preferred. (C) Pooling these edges together results in a curved segment with the correct border-ownership assignment. (D) After this computation is carried out in the local neighborhood, border ownership is largely but not fully correct. We can improve it by using the same convexity assumption over larger areas (e.g., over the entire image). (E) Global border-ownership computation results in a correct assignment of all segments. (F) With pooling, two separate surfaces emerge. Note that the blue one is missing a boundary at the intersection with the yellow object. This implies that the blue object is partially occluded by the yellow one. (G) Using this information, a correct local depth ordering is established and the missing piece of the blue object is interpolated.