next up previous
Next: Results Up: A Worked-Out Example: Depth-Perception Previous: Coherence-Detection

Network-Structure

Disparity estimators in real biological networks can and will vary in various properties, notably in the separation of the center of their receptive fields in the left and right eye. Other possible parameters of interest are the spatial orientation, the scale or the phase of the Gabor filter patches.

Figure 8: The main structure of the stereo vision network. Disparity estimators are arranged in horizontal sheets responding to a common depth. Units looking into common view directions are collected together, into vertically aligned disparity stacks. Within these stacks, coherence detection takes place, facilitated by appropriate weak coupling between the units in a stack.
\resizebox* {0.7\columnwidth}{!}{\includegraphics{figs/network/network.eps}}

In the simplified network used here, the orientation of the Gabor filter patches is fixed to the vertical, and three to four different spatial scales are used. At a single scale, image data is fed diagonally into layers of identical disparity units. This creates layers of units having a common and fixed separation of receptive fields in the left and right eye (Fig. 8). Disparity units stacked vertically above each other sample space in a common view direction, and it is here where the coherence detection scheme sets in: all units with a common view direction are connected with each other through weak synaptic links. This coupling creates the vertical disparity stacks marked by the rectangular outlines in Fig 8. To simplify the simulations, disparity units operating at different scales are not treated separately, but are simply included in the appropriate disparity stacks at the finest spatial resolution.


next up previous
Next: Results Up: A Worked-Out Example: Depth-Perception Previous: Coherence-Detection

2000-11-20