next up previous
Next: The Coherence Network Up: The Input Layers Previous: Complex Cells

Disparity Estimators

The output of any complex cell depends also on the local image contrast and has to be normalized in order to achieve contrast-independent performance. The necessary contrast normalization, which might be realized in cortex by shunting inhibition [40,46,47], was modeled here simply by dividing the signals of complex cells responding to left and right image motion with the response of a matching complex cell measuring local contrast. Following Adelson, we construct disparity estimators out of the response of three complex cells by computing

d_{(i,j)}(s)=\frac{c_{(i,j)}(+\pi /4,s)-c_{(i,j)}(-\pi /4,s)}{c_{(i,j)}(0,s)}\, .\end{displaymath}

In effect, this gives a set of noisy local disparity estimates \( d_{(i,j)}(s) \) in each view direction \( (i,j) \). The shift parameter \( s \) becomes a this point merely an index, numbering different disparity units in a single disparity stack \( \mathcal{D}(i,j) \) [45].