Dual Processing of Visual Motion Reduces Smearing, Delay and Noise, But Yields the ‘Wavy Edge’ and ‘Window-Shift’ Illusions
Early stages of visual motion processing are modelled by two networks, each consisting of velocity-tuned cells. One network extracts the (possibly multi-valued) velocity-field, while the other extracts the patterns which move. The ‘channel-coded’ output from the velocity-net selects those co-local units in the pattern-net which are tuned to the local velocity, so preventing motion-smear and allowing compensation of target positions for unavoidable processing delays. Under special conditions, however, the model creates some artefacts which correspond well with known visual illusions.
KeywordsManifold Assimilation Convolution
Unable to display preview. Download preview PDF.