EXISTING TECHNIQUES OF VISUAL PROCESSING - Video Object Extraction and Representation: Theory and Applications - page 35

Existing Techniques of Visual Processing 35

Figure 2.5. Two fundamental ambiguities of the optical flow constraint are demon-
strated where position of the objects the next frame is shown by the dashed boxes. a)
The Aperture Problem: If we try to determine the movement of the middle of a side,
the correspondence of a block in the next frame is ambiguous. However, if we zoom
out and take a larger context, the movement can be clearly resolved. b) Correspon-
dence Problem: The movement of two identical blocks upward has two explanations:
either both blocks moved upward as on the left, or they diagonally switched places
between frames as on the right.

smoothness, to Eq. 2.1. The effect of changing lighting conditions cannot
be avoided and these lighting conditions must be caught before optical
flow processing. In our work, most of our test sequences do not have
such difficult lighting conditions.
However, even with these assumptions, there exists two inherent prob-
lems with the optical flow assumption: the aperture problem and the
correspondence problem (to be defined, shortly). These problems are
related to the fact that although projected motion is considered a low-
level feature, motion contains much high-level information such as object
membership and object motion.
As shown in Figure 2.5a, the aperture problem is caused by the lack of
the contextual information to resolve visual ambiguities. For example,
consider a section of the homogeneously-filled square moving from one
image to the next. If we consider a macroblock from only one side of
the square in isolation, its motion is ambiguous. Without knowledge of
the corners, it is difficult to predict the motion of the sides. Without
knowledge of the sides, it is difficult to predict the bulk of the square.
Both are cases of the aperture problem. Note that if there was a vertical