!Converted with LaTeX2HTML 95.1 (Fri Jan 20 1995) by Nikos Drakos (firstname.lastname@example.org), CBLU, University of Leeds >
During motion, visual objects undergo substantial changes in appearance. They change size, shape, and position with respect to the background (Figure 1). They even occasionally disappear behind other objects (Figure 1c) and reappear in a new position (Figure 1d).
Figure 1: Changes of appearance during an occlusion event. (a) Image frame containing two objects. Circle is moving toward the right. (b) Circle becomes partially occluded, and its image shape changes. (c) Circle becomes completely occluded. (d) Circle begins to emerge on the other side of the occluder, and its shape is again changed.
Our visual systems make highly effective use of these changes to deduce the depth relations among objects. For instance, the occlusion sequence in Figure 1b--d) is typically judged to indicate that the rectangle is nearer in depth than the circle. Evidence from psychophysics [6,12,13,15,16,17] and neurophysiology  suggests that the process of determining relative depth from occlusion events operates at an early stage of visual processing.
Marshall  describes evidence that suggests that the same early processing mechanisms maintain a representation of temporarily occluded objects for some amount of time after they have disappeared behind an occluder, and that these representations of invisible objects interact with other object representations, in much the same manner as do representations of visible objects. For example, Shimojo, Silverman, & Nakayama  describe a way in which our visual mechanisms for processing motion information and stereo depth information interact despite the temporary occlusion of a moving object in one or both eyes.
Below we describe how a visual system can learn to detect and represent depth relations, after a period of exposure to occlusion and disocclusion events. We use a self-organizing neural network model that exploits the visual changes that occur at occlusion boundaries to form a mechanism for detecting and representing relative depth information. The network's learning is governed by a new set of learning and activation rules.