a
«D *9 4.2 7c CO CX CX cr 5
ging problem in photogrammetry. The human operator does it with great
ease, and the performance of computer methods will always implicitly be
compared with that of human operators.
The problem of stereo matching can easily be formulated as a three- stage
process: firstly, a pixel location that corresponds to a point of the surface
of the object is chosen in one of the images; secondly, the corresponding
(conjugate) pixel in the other image (or in case of multiple coverage in all
other images) is located; and finally the three- dimensional position of the
point is determined. Since pixel locations are in fact image coordinates
the latter step can rely on the well- established methods of analytical
photogrammetry.
Traditionally, in photogrammetry the preferred methods for stereo
matching are those of area correlation. The assumption is that the
conjugent points can be found by correlating the image function of the two
images. These methods produce good results in many cases; in others,
however, they fail. It was always hoped that further refinements of the
correlation technique would solve the remaining problems.
During the last ten years, a new theory about the human visual system has
emerged and computer models have been developed to test the theory
(see, e.g., /3/,/6/,/7/). One of the interesting conclusions is the fact that
the human visual system does not find the corresponding points by
matching raw intensity values (gray levels), rather it matches abrupt
changes in the image function which often reflect physica! boundaries in
the object space. When accepting the human visual system as superior to
any automated system the message becomes plain: the area correlation
and its many variants is the wrong method.
Stereomatching in the human visual system
Vision is our most impressive sense. What we do without conscious effort
is a massive information process that has not been well understood until
only recently. "Seeing' seems easy and straightforward so that one
drastically underestimates the problem. It turns out that for a computer to
perform the simplest act of vision requires millions of multiplications.
- 603 -