The stereo
correspondence problem

To perform successful matching of primitives across two images requires the viewer to overcome the stereo correspondence problem. To gain a better understanding of the problem refer to Figure 2, below.

Figure 2:   The stereo correspondence problem

Each eye/camera views four image primitives (dots). The problem then is, which dots in the left eye correspond to which dots in the right eye ? The 9 dots represent all the possible matches that could be made, the black dots are the correct matches and the rest are incorrect, (referred to as either ``false targets'' or ``ghosts'').

Confronted by these 9 possible matches, we ourselves are capable (in this instance) of making the 3 correct matches. The interesting thing about using the dot example is that no high level information or cues are presented to help the viewer in matching. This led researchers to believe that stereo matching is performed early in the human visual process, it is assumed to be a low level operation. Only after the two images are matched is any attempt made to understand what is actually being viewed.

The depth is computed from similar triangles in Figure 3, below.

 Figure 3
Figure 3:   Simple camera geometry

[ Feature detection | Random dot stereograms ]

Comments to: Sarah Price at ICBL.