The generalized maximum likelihood (GML) algorithm is a gradient-based iterative algorithm for frame-to-frame motion estimation. This algorithm tends toward the maximum likelihood estimates of the Karhunen-Loève expansion coefficients of the motion field. The GML algorithm requires the covariance function matrix as a priori knowledge. Determination of the actual motion covariance in a practical situation is a difficult problem; the problem is approached by assuming that the motion vector is modeled by a separable stationary Markov-2 field. Using this model, we relate and compare the GML algorithm to another well-known motion estimator reported by Netravali and Robbins. Simulation experiments are presented that indicate the improvement of the GML algorithm over Netravali's scheme.