Analysis of bias in gradient-based optical-flow estimation

Accurate gradient-based optical-flow estimation depends on accurate partial derivatives that are generally approximated by the use of finite-differencing convolution kernels. The consequent error in the first derivative estimate is approximately proportional to the third derivative of the input signal and leads to systematic errors in the optical-flow estimates. Simulations indicate that these errors tend to dominate other error sources, such as broad-band noise, unless the system is carefully tuned. The result suggests that the high-frequency attenuation of the finite-differencing kernel imposes a resolution limit on the estimated optical-flow field