Motion Perception and Mid-Level Vision

Note: the phenomena described in this chapter are very difficult to understand without viewing the moving stimuli. The reader is urged to view the demos when reading the chapter, at: 2 Like many aspects of vision, motion perception begins with a massive array of local measurements performed by neurons in area V1. Each receptive field covers a small piece of the visual world, and as a result suffers from an ambiguity known as the aperture problem, illustrated in Figure 1. A moving contour, locally observed, is consistent with a family of possible motions (Wallach, 1935; Adelson and Movshon, 1982). This ambiguity is geometric in origin-motion parallel to the contour cannot be detected, as changes to this component of the motion do not change the images observed through the aperture. Only the component of the velocity orthogonal to the contour orientation can be measured, and as a result the actual velocity could be any of an infinite family of motions lying along a line in velocity space, as indicated in Figure 1. This ambiguity depends on the contour in question being straight, but smoothly curved contours are approximately straight when viewed locally, and the aperture problem is thus widespread. The upshot is that most local measurements made in the early stages of vision constrain object velocities but do not narrow them down to a single value; further analysis is necessary to yield the motions that we perceive. It is possible to resolve the ambiguity of local measurements by combining information across space, as shown in Figure 2. The motion of 2-D features, such as the corner marked 2, is unambiguous, and can be combined with the contour information to provide a consistent velocity estimate. On the other hand, some 2-D features are the result of occlusion, such as the T-junction (marked 3) that occurs where the two squares of Figure 2(a) overlap. The motion of such features is spurious and does not correspond to the motion of any single physical object; in Figure 2 the two squares move left and right but the T-junction moves down. Distinguishing spurious features from real ones requires the use of form information, as the motion generated by such features does not in itself distinguish them. An alternate way of extracting 2-D motion is to combine the ambiguous information from different contours of the same object, as shown in Figure 2(c). In velocity space, the constraints …