It has been claimed that inherently sequential approaches to early vision, such as cooperative algorithms, are biologically implausible. The relatively low processing speed of neurons, together with the very rapid response time of biological stereo segmentation, argue against an iterative solution to stereo segmentation in particular, and to sensory information processing in general. To perform in real time, the biological solution should consist of a one-shot, local operator. Such an operator is implemented by a special-purpose local neural network. Stereoscopic vision is used here as a special case of this proposed general principle. Evidence is given showing that stereo positional resolution is coarser than visual acuity by roughly an order of magnitude. It is concluded that this is the result of a submodule of stereo vision being processed in windows and not pointwise. It is suggested that these windows consist of interlaced left and right images (similar to the ocular dominance column format of primate visual cortex), and that these windows are operated upon by a cepstral operator.
|State||Published - 1987|