An adaptive multi-scale approach for estimating optical flow: computational theory and physiological implementation

The accuracy of optical flow estimation depends on the spatio-temporal discretization used in the computation. The authors propose an adaptive multiscale method, where the discretization scale is chosen locally according to an estimate of the relative error in the velocity measurements. They show that their coarse-to-fine method provides substantially better results of optical flow than conventional algorithms. The authors map this multiscale strategy onto their model of motion computation in primate area MT. The model consists of two stages: (1) local velocities are measured across multiple spatio-temporal channels, while (2) the optical flow field is computed by a network of direction-selective neurons at multiple spatial resolutions. Their model neurons show the same nonclassical receptive field properties as Allman's type I MT neurons and lead to a novel interpretation of some aspect of the motion capture illusion.<>