A systematic analysis of a V1–MT neural model for motion estimation

A neural feed-forward model composed of two layers that mimic the V1–MT primary motion pathway, derived from previous works by Heeger and Simoncelli, is proposed and analyzed. Essential aspects of the model are highlighted and comparatively analyzed to point out how realistic neural responses can be efficiently and effectively used for optic flow estimation if properly combined at a population level. First, different profiles of the spatio-temporal V1 receptive fields are compared, both in terms of their properties in the frequency domain, and in terms of their responses to random dots and plaid stimuli. Then, a pooling stage at the MT level, which combines the afferent V1 responses, is modeled to obtain a population of pattern cells that encodes the local velocities of the visual stimuli. Finally, a decoding stage allows us to combine MT activities in order to compute optic flow. A systematic validation of the model is performed by computing the optic flow on synthetic and standard benchmark sequences with ground truth flow available. The average angular errors and the end-point errors on the resulting estimates allow us to quantitatively compare the different spatio-temporal profiles and the choices of the model׳s parameters, and to assess the validity and effectiveness of the approach in realistic situations.